{
  "schemaVersion": 1,
  "deviceProperties": [
    {
      "id": 0, "name": "AMD Instinct MI300X VF", "totalGlobalMem": 205571227648,
      "computeMajor": 9, "computeMinor": 4,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 64,
      "sharedMemPerBlock": 65536, "numSms": 304
    , "maxSharedMemoryPerMultiProcessor": 65536
    }
  ],
  "roctracer_version": 4.1,
  "hip_runtime_version": 60342131,
  "record_shapes": 1,
  "hip_driver_version": 60342131,
  "with_stack": 1,
  "trace_id": "185F7E4226ED4DA7AAF624E75FFB4F2D",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740836874.147, "dur": 202.853,
    "args": {
      "External id": 1,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740836890.595, "dur": 150.269,
    "args": {
      "External id": 2,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740837049.355, "dur": 27.102,
    "args": {
      "External id": 3,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740837103.798, "dur": 0.443,
    "args": {
      "External id": 4,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740837135.102, "dur": 37858.246,
    "args": {
      "External id": 5,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740837193.024, "dur": 20.237,
    "args": {
      "External id": 6,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740837207.728, "dur": 2.081,
    "args": {
      "External id": 7,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740837215.718, "dur": 1.818,
    "args": {
      "External id": 8,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740837216.782, "dur": 0.335,
    "args": {
      "External id": 9,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740837231.550, "dur": 7772.904,
    "args": {
      "External id": 10,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740845034.558, "dur": 5747.677,
    "args": {
      "External id": 11,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740850816.995, "dur": 98.302,
    "args": {
      "External id": 12,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740850939.840, "dur": 50.262,
    "args": {
      "External id": 13,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740851005.063, "dur": 10129.207,
    "args": {
      "External id": 14,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740861192.229, "dur": 13639.125,
    "args": {
      "External id": 15,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740861203.005, "dur": 0.552,
    "args": {
      "External id": 16,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740874858.712, "dur": 33.705,
    "args": {
      "External id": 17,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740874895.993, "dur": 79.413,
    "args": {
      "External id": 18,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875013.922, "dur": 29.800,
    "args": {
      "External id": 19,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875020.384, "dur": 4.313,
    "args": {
      "External id": 20,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875026.513, "dur": 16.853,
    "args": {
      "External id": 21,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875062.558, "dur": 0.514,
    "args": {
      "External id": 22,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875086.147, "dur": 164.940,
    "args": {
      "External id": 23,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875119.839, "dur": 10.172,
    "args": {
      "External id": 24,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875125.618, "dur": 1.725,
    "args": {
      "External id": 25,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875132.138, "dur": 1.750,
    "args": {
      "External id": 26,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 25
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875133.125, "dur": 0.338,
    "args": {
      "External id": 27,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875142.082, "dur": 14.330,
    "args": {
      "External id": 28,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875162.531, "dur": 11.005,
    "args": {
      "External id": 29,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875178.370, "dur": 6.627,
    "args": {
      "External id": 30,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 29
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875188.271, "dur": 10.448,
    "args": {
      "External id": 31,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875202.269, "dur": 9.415,
    "args": {
      "External id": 32,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875216.818, "dur": 11.316,
    "args": {
      "External id": 33,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875218.518, "dur": 0.203,
    "args": {
      "External id": 34,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875230.661, "dur": 7.542,
    "args": {
      "External id": 35,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875239.480, "dur": 6.290,
    "args": {
      "External id": 36,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875257.197, "dur": 11.940,
    "args": {
      "External id": 37,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875258.474, "dur": 1.850,
    "args": {
      "External id": 38,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875260.938, "dur": 7.950,
    "args": {
      "External id": 39,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875276.176, "dur": 0.141,
    "args": {
      "External id": 40,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 39
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875284.406, "dur": 86.634,
    "args": {
      "External id": 41,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875294.383, "dur": 2.373,
    "args": {
      "External id": 42,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875295.716, "dur": 0.472,
    "args": {
      "External id": 43,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875298.013, "dur": 1.376,
    "args": {
      "External id": 44,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875298.785, "dur": 0.282,
    "args": {
      "External id": 45,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875302.284, "dur": 6.723,
    "args": {
      "External id": 46,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875311.766, "dur": 6.843,
    "args": {
      "External id": 47,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875321.290, "dur": 5.797,
    "args": {
      "External id": 48,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875328.673, "dur": 6.370,
    "args": {
      "External id": 49,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875337.153, "dur": 6.494,
    "args": {
      "External id": 50,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 49
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875346.237, "dur": 6.643,
    "args": {
      "External id": 51,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875347.018, "dur": 0.099,
    "args": {
      "External id": 52,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875354.754, "dur": 5.568,
    "args": {
      "External id": 53,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875361.335, "dur": 5.746,
    "args": {
      "External id": 54,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875375.689, "dur": 8.434,
    "args": {
      "External id": 55,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875376.683, "dur": 1.349,
    "args": {
      "External id": 56,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875378.504, "dur": 5.429,
    "args": {
      "External id": 57,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875389.784, "dur": 0.140,
    "args": {
      "External id": 58,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875395.685, "dur": 76.535,
    "args": {
      "External id": 59,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875403.292, "dur": 2.030,
    "args": {
      "External id": 60,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 59
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875404.448, "dur": 0.368,
    "args": {
      "External id": 61,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875406.646, "dur": 1.418,
    "args": {
      "External id": 62,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875407.446, "dur": 0.283,
    "args": {
      "External id": 63,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875410.391, "dur": 6.266,
    "args": {
      "External id": 64,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875418.872, "dur": 5.574,
    "args": {
      "External id": 65,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875426.588, "dur": 5.351,
    "args": {
      "External id": 66,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875433.288, "dur": 5.672,
    "args": {
      "External id": 67,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875440.644, "dur": 5.836,
    "args": {
      "External id": 68,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875448.823, "dur": 6.636,
    "args": {
      "External id": 69,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875449.423, "dur": 0.115,
    "args": {
      "External id": 70,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 69
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875457.053, "dur": 5.872,
    "args": {
      "External id": 71,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875463.916, "dur": 4.960,
    "args": {
      "External id": 72,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875476.147, "dur": 7.796,
    "args": {
      "External id": 73,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875477.042, "dur": 1.343,
    "args": {
      "External id": 74,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875478.807, "dur": 4.953,
    "args": {
      "External id": 75,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 74
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875489.253, "dur": 0.143,
    "args": {
      "External id": 76,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875494.835, "dur": 74.671,
    "args": {
      "External id": 77,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875502.607, "dur": 1.908,
    "args": {
      "External id": 78,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875503.675, "dur": 0.355,
    "args": {
      "External id": 79,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875505.806, "dur": 1.410,
    "args": {
      "External id": 80,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875506.628, "dur": 0.255,
    "args": {
      "External id": 81,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875509.215, "dur": 5.514,
    "args": {
      "External id": 82,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875516.681, "dur": 5.643,
    "args": {
      "External id": 83,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875524.508, "dur": 5.915,
    "args": {
      "External id": 84,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875531.745, "dur": 5.784,
    "args": {
      "External id": 85,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875539.126, "dur": 5.466,
    "args": {
      "External id": 86,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875546.762, "dur": 6.053,
    "args": {
      "External id": 87,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875547.326, "dur": 0.140,
    "args": {
      "External id": 88,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875554.428, "dur": 5.435,
    "args": {
      "External id": 89,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875560.956, "dur": 5.396,
    "args": {
      "External id": 90,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 89
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875573.097, "dur": 7.979,
    "args": {
      "External id": 91,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875574.189, "dur": 1.269,
    "args": {
      "External id": 92,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875575.950, "dur": 4.983,
    "args": {
      "External id": 93,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875585.889, "dur": 0.130,
    "args": {
      "External id": 94,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 93
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875591.205, "dur": 73.254,
    "args": {
      "External id": 95,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875597.967, "dur": 1.839,
    "args": {
      "External id": 96,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875599.016, "dur": 0.306,
    "args": {
      "External id": 97,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875601.078, "dur": 1.309,
    "args": {
      "External id": 98,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875601.832, "dur": 0.252,
    "args": {
      "External id": 99,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875604.394, "dur": 5.928,
    "args": {
      "External id": 100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875612.179, "dur": 5.526,
    "args": {
      "External id": 101,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875619.648, "dur": 5.469,
    "args": {
      "External id": 102,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875626.295, "dur": 5.535,
    "args": {
      "External id": 103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875633.427, "dur": 5.697,
    "args": {
      "External id": 104,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875641.271, "dur": 6.340,
    "args": {
      "External id": 105,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875641.837, "dur": 0.114,
    "args": {
      "External id": 106,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875649.275, "dur": 5.747,
    "args": {
      "External id": 107,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875655.987, "dur": 5.463,
    "args": {
      "External id": 108,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875667.643, "dur": 7.746,
    "args": {
      "External id": 109,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875668.451, "dur": 1.312,
    "args": {
      "External id": 110,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875670.202, "dur": 5.051,
    "args": {
      "External id": 111,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875680.012, "dur": 0.151,
    "args": {
      "External id": 112,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875685.003, "dur": 88.353,
    "args": {
      "External id": 113,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875691.469, "dur": 1.943,
    "args": {
      "External id": 114,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875692.590, "dur": 0.320,
    "args": {
      "External id": 115,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875694.650, "dur": 1.424,
    "args": {
      "External id": 116,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875695.413, "dur": 0.352,
    "args": {
      "External id": 117,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875697.971, "dur": 5.930,
    "args": {
      "External id": 118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875705.705, "dur": 20.437,
    "args": {
      "External id": 119,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875728.111, "dur": 5.616,
    "args": {
      "External id": 120,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875734.971, "dur": 5.547,
    "args": {
      "External id": 121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875742.127, "dur": 6.031,
    "args": {
      "External id": 122,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875750.252, "dur": 6.163,
    "args": {
      "External id": 123,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875750.891, "dur": 0.099,
    "args": {
      "External id": 124,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875757.867, "dur": 5.597,
    "args": {
      "External id": 125,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875764.439, "dur": 6.081,
    "args": {
      "External id": 126,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875776.499, "dur": 29.238,
    "args": {
      "External id": 127,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875777.468, "dur": 1.237,
    "args": {
      "External id": 128,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875779.232, "dur": 26.331,
    "args": {
      "External id": 129,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875810.439, "dur": 0.147,
    "args": {
      "External id": 130,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875815.606, "dur": 75.508,
    "args": {
      "External id": 131,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875822.419, "dur": 2.070,
    "args": {
      "External id": 132,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875823.631, "dur": 0.401,
    "args": {
      "External id": 133,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875825.754, "dur": 1.427,
    "args": {
      "External id": 134,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875826.598, "dur": 0.269,
    "args": {
      "External id": 135,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875829.096, "dur": 6.281,
    "args": {
      "External id": 136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875837.409, "dur": 6.077,
    "args": {
      "External id": 137,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875845.535, "dur": 5.657,
    "args": {
      "External id": 138,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875852.544, "dur": 5.411,
    "args": {
      "External id": 139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875859.494, "dur": 5.880,
    "args": {
      "External id": 140,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875867.454, "dur": 6.237,
    "args": {
      "External id": 141,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875868.093, "dur": 0.088,
    "args": {
      "External id": 142,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875875.245, "dur": 6.204,
    "args": {
      "External id": 143,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875882.376, "dur": 5.680,
    "args": {
      "External id": 144,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875894.446, "dur": 7.960,
    "args": {
      "External id": 145,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875895.484, "dur": 1.282,
    "args": {
      "External id": 146,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875897.280, "dur": 4.963,
    "args": {
      "External id": 147,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875906.785, "dur": 0.154,
    "args": {
      "External id": 148,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875911.609, "dur": 75.124,
    "args": {
      "External id": 149,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875917.960, "dur": 1.899,
    "args": {
      "External id": 150,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875919.027, "dur": 0.353,
    "args": {
      "External id": 151,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740875921.186, "dur": 1.406,
    "args": {
      "External id": 152,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875922.017, "dur": 0.259,
    "args": {
      "External id": 153,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740875924.402, "dur": 6.387,
    "args": {
      "External id": 154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875932.712, "dur": 6.264,
    "args": {
      "External id": 155,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875941.049, "dur": 5.422,
    "args": {
      "External id": 156,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740875947.834, "dur": 6.443,
    "args": {
      "External id": 157,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740875955.954, "dur": 6.244,
    "args": {
      "External id": 158,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740875964.351, "dur": 5.955,
    "args": {
      "External id": 159,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740875964.982, "dur": 0.091,
    "args": {
      "External id": 160,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740875971.912, "dur": 5.463,
    "args": {
      "External id": 161,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740875978.328, "dur": 5.430,
    "args": {
      "External id": 162,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740875989.901, "dur": 7.654,
    "args": {
      "External id": 163,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740875990.805, "dur": 1.194,
    "args": {
      "External id": 164,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740875992.475, "dur": 4.926,
    "args": {
      "External id": 165,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876001.858, "dur": 0.170,
    "args": {
      "External id": 166,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876006.624, "dur": 74.634,
    "args": {
      "External id": 167,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876012.691, "dur": 1.868,
    "args": {
      "External id": 168,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876013.740, "dur": 0.372,
    "args": {
      "External id": 169,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876015.794, "dur": 1.330,
    "args": {
      "External id": 170,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876016.591, "dur": 0.249,
    "args": {
      "External id": 171,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876019.069, "dur": 6.519,
    "args": {
      "External id": 172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876027.626, "dur": 5.958,
    "args": {
      "External id": 173,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876035.405, "dur": 5.463,
    "args": {
      "External id": 174,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876042.204, "dur": 5.228,
    "args": {
      "External id": 175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876049.009, "dur": 5.565,
    "args": {
      "External id": 176,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876056.645, "dur": 6.000,
    "args": {
      "External id": 177,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876057.201, "dur": 0.122,
    "args": {
      "External id": 178,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876064.238, "dur": 6.526,
    "args": {
      "External id": 179,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876071.790, "dur": 6.528,
    "args": {
      "External id": 180,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876084.611, "dur": 10.804,
    "args": {
      "External id": 181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876085.475, "dur": 1.238,
    "args": {
      "External id": 182,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876087.221, "dur": 8.035,
    "args": {
      "External id": 183,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876099.826, "dur": 0.159,
    "args": {
      "External id": 184,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876104.509, "dur": 77.708,
    "args": {
      "External id": 185,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876110.773, "dur": 1.849,
    "args": {
      "External id": 186,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876111.833, "dur": 0.352,
    "args": {
      "External id": 187,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876113.839, "dur": 1.329,
    "args": {
      "External id": 188,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876114.621, "dur": 0.246,
    "args": {
      "External id": 189,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876117.090, "dur": 6.725,
    "args": {
      "External id": 190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876125.889, "dur": 5.752,
    "args": {
      "External id": 191,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876133.498, "dur": 6.178,
    "args": {
      "External id": 192,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876141.019, "dur": 6.140,
    "args": {
      "External id": 193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876148.736, "dur": 6.980,
    "args": {
      "External id": 194,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876157.892, "dur": 6.763,
    "args": {
      "External id": 195,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876158.429, "dur": 0.112,
    "args": {
      "External id": 196,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876166.181, "dur": 6.285,
    "args": {
      "External id": 197,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876173.401, "dur": 5.933,
    "args": {
      "External id": 198,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876185.526, "dur": 7.712,
    "args": {
      "External id": 199,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876186.408, "dur": 1.122,
    "args": {
      "External id": 200,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876188.097, "dur": 4.978,
    "args": {
      "External id": 201,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876198.655, "dur": 0.174,
    "args": {
      "External id": 202,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876203.600, "dur": 82.191,
    "args": {
      "External id": 203,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876210.511, "dur": 2.043,
    "args": {
      "External id": 204,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876211.682, "dur": 0.424,
    "args": {
      "External id": 205,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876213.842, "dur": 1.402,
    "args": {
      "External id": 206,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876214.704, "dur": 0.242,
    "args": {
      "External id": 207,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876217.194, "dur": 5.822,
    "args": {
      "External id": 208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876224.926, "dur": 6.511,
    "args": {
      "External id": 209,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876233.538, "dur": 7.383,
    "args": {
      "External id": 210,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876242.293, "dur": 7.538,
    "args": {
      "External id": 211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876251.403, "dur": 7.132,
    "args": {
      "External id": 212,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876260.813, "dur": 6.971,
    "args": {
      "External id": 213,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876261.365, "dur": 0.129,
    "args": {
      "External id": 214,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876269.359, "dur": 6.332,
    "args": {
      "External id": 215,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876276.709, "dur": 5.958,
    "args": {
      "External id": 216,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876289.007, "dur": 9.061,
    "args": {
      "External id": 217,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876289.891, "dur": 1.189,
    "args": {
      "External id": 218,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876291.592, "dur": 6.320,
    "args": {
      "External id": 219,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876302.710, "dur": 0.180,
    "args": {
      "External id": 220,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876307.372, "dur": 80.420,
    "args": {
      "External id": 221,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876313.750, "dur": 2.003,
    "args": {
      "External id": 222,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876314.857, "dur": 0.437,
    "args": {
      "External id": 223,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876317.069, "dur": 1.407,
    "args": {
      "External id": 224,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876317.897, "dur": 0.262,
    "args": {
      "External id": 225,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876320.408, "dur": 6.874,
    "args": {
      "External id": 226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876329.139, "dur": 6.504,
    "args": {
      "External id": 227,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876337.884, "dur": 6.140,
    "args": {
      "External id": 228,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876345.323, "dur": 5.779,
    "args": {
      "External id": 229,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876352.664, "dur": 7.578,
    "args": {
      "External id": 230,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876362.537, "dur": 6.781,
    "args": {
      "External id": 231,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876363.102, "dur": 0.110,
    "args": {
      "External id": 232,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876370.841, "dur": 6.249,
    "args": {
      "External id": 233,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876378.994, "dur": 5.871,
    "args": {
      "External id": 234,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876391.046, "dur": 8.331,
    "args": {
      "External id": 235,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876391.934, "dur": 1.264,
    "args": {
      "External id": 236,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876393.715, "dur": 5.525,
    "args": {
      "External id": 237,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876403.684, "dur": 0.200,
    "args": {
      "External id": 238,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876408.255, "dur": 78.831,
    "args": {
      "External id": 239,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876414.424, "dur": 1.988,
    "args": {
      "External id": 240,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876415.639, "dur": 0.344,
    "args": {
      "External id": 241,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876417.691, "dur": 1.474,
    "args": {
      "External id": 242,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876418.487, "dur": 0.372,
    "args": {
      "External id": 243,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876421.073, "dur": 6.295,
    "args": {
      "External id": 244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876429.180, "dur": 6.593,
    "args": {
      "External id": 245,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876437.695, "dur": 7.650,
    "args": {
      "External id": 246,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876446.708, "dur": 5.844,
    "args": {
      "External id": 247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876454.091, "dur": 6.473,
    "args": {
      "External id": 248,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876462.683, "dur": 6.509,
    "args": {
      "External id": 249,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876463.229, "dur": 0.108,
    "args": {
      "External id": 250,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876470.771, "dur": 6.305,
    "args": {
      "External id": 251,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876478.101, "dur": 6.096,
    "args": {
      "External id": 252,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876490.149, "dur": 7.568,
    "args": {
      "External id": 253,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876491.048, "dur": 1.182,
    "args": {
      "External id": 254,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876492.722, "dur": 4.847,
    "args": {
      "External id": 255,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876502.037, "dur": 0.169,
    "args": {
      "External id": 256,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876506.785, "dur": 75.983,
    "args": {
      "External id": 257,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876512.737, "dur": 1.973,
    "args": {
      "External id": 258,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876513.882, "dur": 0.379,
    "args": {
      "External id": 259,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876516.024, "dur": 1.679,
    "args": {
      "External id": 260,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876516.826, "dur": 0.551,
    "args": {
      "External id": 261,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876519.651, "dur": 6.729,
    "args": {
      "External id": 262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876528.217, "dur": 6.310,
    "args": {
      "External id": 263,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876536.479, "dur": 5.744,
    "args": {
      "External id": 264,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876543.504, "dur": 6.198,
    "args": {
      "External id": 265,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876551.146, "dur": 6.228,
    "args": {
      "External id": 266,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876559.488, "dur": 5.822,
    "args": {
      "External id": 267,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876560.013, "dur": 0.101,
    "args": {
      "External id": 268,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876566.914, "dur": 5.997,
    "args": {
      "External id": 269,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876573.910, "dur": 5.959,
    "args": {
      "External id": 270,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876585.866, "dur": 8.700,
    "args": {
      "External id": 271,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876586.754, "dur": 1.185,
    "args": {
      "External id": 272,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876588.417, "dur": 5.994,
    "args": {
      "External id": 273,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876598.822, "dur": 0.139,
    "args": {
      "External id": 274,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876603.459, "dur": 78.715,
    "args": {
      "External id": 275,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876609.702, "dur": 1.837,
    "args": {
      "External id": 276,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876610.765, "dur": 0.333,
    "args": {
      "External id": 277,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876612.773, "dur": 1.426,
    "args": {
      "External id": 278,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876613.622, "dur": 0.257,
    "args": {
      "External id": 279,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876616.123, "dur": 6.653,
    "args": {
      "External id": 280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876624.838, "dur": 6.411,
    "args": {
      "External id": 281,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876633.029, "dur": 6.302,
    "args": {
      "External id": 282,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876640.617, "dur": 5.838,
    "args": {
      "External id": 283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876647.910, "dur": 7.222,
    "args": {
      "External id": 284,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876657.152, "dur": 7.184,
    "args": {
      "External id": 285,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876657.782, "dur": 0.085,
    "args": {
      "External id": 286,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876665.890, "dur": 6.377,
    "args": {
      "External id": 287,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876673.210, "dur": 6.073,
    "args": {
      "External id": 288,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876685.282, "dur": 8.524,
    "args": {
      "External id": 289,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876686.144, "dur": 1.166,
    "args": {
      "External id": 290,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876687.791, "dur": 5.883,
    "args": {
      "External id": 291,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876698.180, "dur": 0.150,
    "args": {
      "External id": 292,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876702.682, "dur": 76.109,
    "args": {
      "External id": 293,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876708.740, "dur": 1.980,
    "args": {
      "External id": 294,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876709.865, "dur": 0.381,
    "args": {
      "External id": 295,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876711.964, "dur": 1.407,
    "args": {
      "External id": 296,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876712.780, "dur": 0.275,
    "args": {
      "External id": 297,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876715.290, "dur": 5.693,
    "args": {
      "External id": 298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876722.875, "dur": 5.621,
    "args": {
      "External id": 299,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876730.322, "dur": 5.259,
    "args": {
      "External id": 300,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876736.887, "dur": 7.061,
    "args": {
      "External id": 301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876745.415, "dur": 6.372,
    "args": {
      "External id": 302,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876754.019, "dur": 6.503,
    "args": {
      "External id": 303,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876754.583, "dur": 0.089,
    "args": {
      "External id": 304,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876762.025, "dur": 6.892,
    "args": {
      "External id": 305,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876769.854, "dur": 6.067,
    "args": {
      "External id": 306,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876781.922, "dur": 7.649,
    "args": {
      "External id": 307,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876782.750, "dur": 1.171,
    "args": {
      "External id": 308,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876784.407, "dur": 5.025,
    "args": {
      "External id": 309,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876793.770, "dur": 0.135,
    "args": {
      "External id": 310,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876798.244, "dur": 79.948,
    "args": {
      "External id": 311,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876805.749, "dur": 1.999,
    "args": {
      "External id": 312,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876806.960, "dur": 0.373,
    "args": {
      "External id": 313,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876808.910, "dur": 1.378,
    "args": {
      "External id": 314,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876809.730, "dur": 0.255,
    "args": {
      "External id": 315,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876812.110, "dur": 7.083,
    "args": {
      "External id": 316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876821.017, "dur": 6.475,
    "args": {
      "External id": 317,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876829.421, "dur": 6.117,
    "args": {
      "External id": 318,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876836.822, "dur": 6.339,
    "args": {
      "External id": 319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876844.571, "dur": 6.559,
    "args": {
      "External id": 320,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876853.247, "dur": 6.244,
    "args": {
      "External id": 321,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876853.848, "dur": 0.095,
    "args": {
      "External id": 322,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876860.958, "dur": 6.555,
    "args": {
      "External id": 323,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876868.494, "dur": 6.944,
    "args": {
      "External id": 324,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740876881.421, "dur": 8.265,
    "args": {
      "External id": 325,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876882.271, "dur": 1.129,
    "args": {
      "External id": 326,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740876883.909, "dur": 5.630,
    "args": {
      "External id": 327,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876894.007, "dur": 0.174,
    "args": {
      "External id": 328,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876898.449, "dur": 108.849,
    "args": {
      "External id": 329,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876904.442, "dur": 1.854,
    "args": {
      "External id": 330,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876905.545, "dur": 0.337,
    "args": {
      "External id": 331,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740876907.453, "dur": 1.385,
    "args": {
      "External id": 332,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740876908.277, "dur": 0.265,
    "args": {
      "External id": 333,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740876910.740, "dur": 33.768,
    "args": {
      "External id": 334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876946.460, "dur": 7.060,
    "args": {
      "External id": 335,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876955.660, "dur": 6.851,
    "args": {
      "External id": 336,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740876963.951, "dur": 7.001,
    "args": {
      "External id": 337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740876972.396, "dur": 7.267,
    "args": {
      "External id": 338,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740876981.794, "dur": 7.131,
    "args": {
      "External id": 339,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740876982.400, "dur": 0.110,
    "args": {
      "External id": 340,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740876990.487, "dur": 6.391,
    "args": {
      "External id": 341,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740876997.797, "dur": 6.593,
    "args": {
      "External id": 342,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877011.025, "dur": 9.360,
    "args": {
      "External id": 343,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877011.944, "dur": 1.212,
    "args": {
      "External id": 344,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877013.666, "dur": 6.583,
    "args": {
      "External id": 345,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877024.751, "dur": 0.138,
    "args": {
      "External id": 346,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877029.734, "dur": 82.769,
    "args": {
      "External id": 347,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877036.027, "dur": 1.970,
    "args": {
      "External id": 348,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877037.176, "dur": 0.374,
    "args": {
      "External id": 349,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877039.243, "dur": 1.382,
    "args": {
      "External id": 350,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877040.058, "dur": 0.260,
    "args": {
      "External id": 351,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877042.675, "dur": 7.191,
    "args": {
      "External id": 352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877051.801, "dur": 8.954,
    "args": {
      "External id": 353,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877062.720, "dur": 5.693,
    "args": {
      "External id": 354,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877069.740, "dur": 5.892,
    "args": {
      "External id": 355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877077.078, "dur": 6.664,
    "args": {
      "External id": 356,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877086.048, "dur": 7.843,
    "args": {
      "External id": 357,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877086.671, "dur": 0.105,
    "args": {
      "External id": 358,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877095.344, "dur": 6.557,
    "args": {
      "External id": 359,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877102.905, "dur": 6.791,
    "args": {
      "External id": 360,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877115.624, "dur": 9.196,
    "args": {
      "External id": 361,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877116.567, "dur": 1.197,
    "args": {
      "External id": 362,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877118.423, "dur": 6.254,
    "args": {
      "External id": 363,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877129.129, "dur": 0.151,
    "args": {
      "External id": 364,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877133.930, "dur": 80.024,
    "args": {
      "External id": 365,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877140.167, "dur": 1.990,
    "args": {
      "External id": 366,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877141.365, "dur": 0.364,
    "args": {
      "External id": 367,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877143.327, "dur": 1.427,
    "args": {
      "External id": 368,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877144.157, "dur": 0.285,
    "args": {
      "External id": 369,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877146.518, "dur": 5.700,
    "args": {
      "External id": 370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877153.964, "dur": 6.594,
    "args": {
      "External id": 371,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877162.383, "dur": 7.259,
    "args": {
      "External id": 372,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877171.031, "dur": 7.170,
    "args": {
      "External id": 373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877179.709, "dur": 7.413,
    "args": {
      "External id": 374,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877189.197, "dur": 6.746,
    "args": {
      "External id": 375,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877189.852, "dur": 0.119,
    "args": {
      "External id": 376,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877197.383, "dur": 6.772,
    "args": {
      "External id": 377,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877205.122, "dur": 6.177,
    "args": {
      "External id": 378,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877217.062, "dur": 7.630,
    "args": {
      "External id": 379,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877217.970, "dur": 1.173,
    "args": {
      "External id": 380,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877219.635, "dur": 4.908,
    "args": {
      "External id": 381,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877228.838, "dur": 0.171,
    "args": {
      "External id": 382,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877233.325, "dur": 80.694,
    "args": {
      "External id": 383,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877239.377, "dur": 1.888,
    "args": {
      "External id": 384,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877240.388, "dur": 0.399,
    "args": {
      "External id": 385,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877242.488, "dur": 1.397,
    "args": {
      "External id": 386,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877243.288, "dur": 0.274,
    "args": {
      "External id": 387,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877245.742, "dur": 7.247,
    "args": {
      "External id": 388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877255.034, "dur": 7.845,
    "args": {
      "External id": 389,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877264.780, "dur": 6.084,
    "args": {
      "External id": 390,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877272.206, "dur": 5.777,
    "args": {
      "External id": 391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877279.468, "dur": 7.287,
    "args": {
      "External id": 392,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877288.831, "dur": 6.148,
    "args": {
      "External id": 393,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877289.494, "dur": 0.142,
    "args": {
      "External id": 394,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877296.487, "dur": 6.410,
    "args": {
      "External id": 395,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877303.854, "dur": 7.447,
    "args": {
      "External id": 396,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877317.117, "dur": 8.707,
    "args": {
      "External id": 397,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877318.137, "dur": 1.190,
    "args": {
      "External id": 398,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877319.842, "dur": 5.832,
    "args": {
      "External id": 399,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877330.144, "dur": 0.167,
    "args": {
      "External id": 400,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877334.556, "dur": 80.703,
    "args": {
      "External id": 401,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877340.481, "dur": 1.893,
    "args": {
      "External id": 402,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877341.571, "dur": 0.371,
    "args": {
      "External id": 403,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877343.576, "dur": 1.397,
    "args": {
      "External id": 404,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877344.392, "dur": 0.265,
    "args": {
      "External id": 405,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877346.760, "dur": 6.611,
    "args": {
      "External id": 406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877355.102, "dur": 6.022,
    "args": {
      "External id": 407,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877362.846, "dur": 7.417,
    "args": {
      "External id": 408,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877371.576, "dur": 6.739,
    "args": {
      "External id": 409,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877379.849, "dur": 7.430,
    "args": {
      "External id": 410,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877389.315, "dur": 7.451,
    "args": {
      "External id": 411,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877389.912, "dur": 0.115,
    "args": {
      "External id": 412,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877398.220, "dur": 7.295,
    "args": {
      "External id": 413,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877406.431, "dur": 6.088,
    "args": {
      "External id": 414,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877418.300, "dur": 8.340,
    "args": {
      "External id": 415,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877419.306, "dur": 1.147,
    "args": {
      "External id": 416,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877421.036, "dur": 5.463,
    "args": {
      "External id": 417,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877430.747, "dur": 0.157,
    "args": {
      "External id": 418,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877435.133, "dur": 78.804,
    "args": {
      "External id": 419,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877440.987, "dur": 2.050,
    "args": {
      "External id": 420,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877442.104, "dur": 0.517,
    "args": {
      "External id": 421,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877444.247, "dur": 2.113,
    "args": {
      "External id": 422,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877445.552, "dur": 0.514,
    "args": {
      "External id": 423,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877448.129, "dur": 7.388,
    "args": {
      "External id": 424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877457.296, "dur": 6.798,
    "args": {
      "External id": 425,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877465.977, "dur": 5.960,
    "args": {
      "External id": 426,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877473.216, "dur": 5.985,
    "args": {
      "External id": 427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877480.732, "dur": 6.188,
    "args": {
      "External id": 428,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877489.044, "dur": 6.621,
    "args": {
      "External id": 429,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877489.579, "dur": 0.108,
    "args": {
      "External id": 430,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877497.068, "dur": 6.962,
    "args": {
      "External id": 431,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877504.982, "dur": 6.278,
    "args": {
      "External id": 432,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877517.137, "dur": 7.976,
    "args": {
      "External id": 433,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877518.023, "dur": 1.233,
    "args": {
      "External id": 434,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877519.755, "dur": 5.206,
    "args": {
      "External id": 435,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877529.464, "dur": 0.169,
    "args": {
      "External id": 436,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877533.926, "dur": 78.622,
    "args": {
      "External id": 437,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877540.043, "dur": 1.936,
    "args": {
      "External id": 438,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877541.141, "dur": 0.402,
    "args": {
      "External id": 439,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877543.108, "dur": 1.422,
    "args": {
      "External id": 440,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877543.917, "dur": 0.266,
    "args": {
      "External id": 441,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877546.451, "dur": 7.522,
    "args": {
      "External id": 442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877555.921, "dur": 6.060,
    "args": {
      "External id": 443,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877563.770, "dur": 5.946,
    "args": {
      "External id": 444,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877571.102, "dur": 6.541,
    "args": {
      "External id": 445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877579.151, "dur": 6.567,
    "args": {
      "External id": 446,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877587.761, "dur": 6.443,
    "args": {
      "External id": 447,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877588.337, "dur": 0.110,
    "args": {
      "External id": 448,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877595.693, "dur": 6.016,
    "args": {
      "External id": 449,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877603.632, "dur": 6.180,
    "args": {
      "External id": 450,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877615.696, "dur": 7.990,
    "args": {
      "External id": 451,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877616.606, "dur": 1.138,
    "args": {
      "External id": 452,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877618.223, "dur": 5.320,
    "args": {
      "External id": 453,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877627.966, "dur": 0.170,
    "args": {
      "External id": 454,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877632.454, "dur": 77.336,
    "args": {
      "External id": 455,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877638.739, "dur": 1.952,
    "args": {
      "External id": 456,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877639.851, "dur": 0.400,
    "args": {
      "External id": 457,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877641.909, "dur": 1.413,
    "args": {
      "External id": 458,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877642.743, "dur": 0.279,
    "args": {
      "External id": 459,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877645.107, "dur": 6.275,
    "args": {
      "External id": 460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877653.229, "dur": 6.990,
    "args": {
      "External id": 461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877662.096, "dur": 6.310,
    "args": {
      "External id": 462,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877669.787, "dur": 5.962,
    "args": {
      "External id": 463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877677.293, "dur": 6.180,
    "args": {
      "External id": 464,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877685.643, "dur": 6.778,
    "args": {
      "External id": 465,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877686.298, "dur": 0.110,
    "args": {
      "External id": 466,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877694.492, "dur": 6.254,
    "args": {
      "External id": 467,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877701.766, "dur": 5.269,
    "args": {
      "External id": 468,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877713.058, "dur": 7.551,
    "args": {
      "External id": 469,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877713.948, "dur": 1.173,
    "args": {
      "External id": 470,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877715.638, "dur": 4.829,
    "args": {
      "External id": 471,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877725.573, "dur": 0.175,
    "args": {
      "External id": 472,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877730.239, "dur": 89.206,
    "args": {
      "External id": 473,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877741.543, "dur": 2.349,
    "args": {
      "External id": 474,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877742.716, "dur": 0.688,
    "args": {
      "External id": 475,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877745.162, "dur": 1.434,
    "args": {
      "External id": 476,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877746.006, "dur": 0.252,
    "args": {
      "External id": 477,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877748.588, "dur": 5.901,
    "args": {
      "External id": 478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877756.410, "dur": 6.756,
    "args": {
      "External id": 479,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877765.132, "dur": 7.053,
    "args": {
      "External id": 480,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877773.498, "dur": 8.174,
    "args": {
      "External id": 481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877783.246, "dur": 8.208,
    "args": {
      "External id": 482,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877793.577, "dur": 6.383,
    "args": {
      "External id": 483,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877794.226, "dur": 0.133,
    "args": {
      "External id": 484,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877801.521, "dur": 6.830,
    "args": {
      "External id": 485,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877809.351, "dur": 7.144,
    "args": {
      "External id": 486,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877822.825, "dur": 8.432,
    "args": {
      "External id": 487,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877823.701, "dur": 1.188,
    "args": {
      "External id": 488,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877825.366, "dur": 5.757,
    "args": {
      "External id": 489,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877835.875, "dur": 0.123,
    "args": {
      "External id": 490,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877840.601, "dur": 84.097,
    "args": {
      "External id": 491,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877846.918, "dur": 1.918,
    "args": {
      "External id": 492,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877847.991, "dur": 0.373,
    "args": {
      "External id": 493,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877850.101, "dur": 1.428,
    "args": {
      "External id": 494,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877850.935, "dur": 0.259,
    "args": {
      "External id": 495,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877853.413, "dur": 5.898,
    "args": {
      "External id": 496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877860.954, "dur": 8.651,
    "args": {
      "External id": 497,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877871.622, "dur": 6.530,
    "args": {
      "External id": 498,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740877879.536, "dur": 5.760,
    "args": {
      "External id": 499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740877886.837, "dur": 9.198,
    "args": {
      "External id": 500,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740877898.176, "dur": 6.822,
    "args": {
      "External id": 501,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877898.904, "dur": 0.113,
    "args": {
      "External id": 502,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740877906.446, "dur": 6.618,
    "args": {
      "External id": 503,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877914.044, "dur": 7.749,
    "args": {
      "External id": 504,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740877927.934, "dur": 8.013,
    "args": {
      "External id": 505,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877928.800, "dur": 1.293,
    "args": {
      "External id": 506,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740877930.590, "dur": 5.210,
    "args": {
      "External id": 507,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740877940.404, "dur": 0.116,
    "args": {
      "External id": 508,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877944.936, "dur": 107.216,
    "args": {
      "External id": 509,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877951.079, "dur": 1.942,
    "args": {
      "External id": 510,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877952.157, "dur": 0.413,
    "args": {
      "External id": 511,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740877954.195, "dur": 28.821,
    "args": {
      "External id": 512,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740877982.309, "dur": 0.383,
    "args": {
      "External id": 513,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740877984.923, "dur": 7.415,
    "args": {
      "External id": 514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740877994.169, "dur": 6.189,
    "args": {
      "External id": 515,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878002.261, "dur": 6.148,
    "args": {
      "External id": 516,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878009.861, "dur": 6.312,
    "args": {
      "External id": 517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878017.712, "dur": 6.989,
    "args": {
      "External id": 518,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878026.848, "dur": 7.275,
    "args": {
      "External id": 519,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878027.587, "dur": 0.102,
    "args": {
      "External id": 520,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878035.627, "dur": 6.402,
    "args": {
      "External id": 521,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878042.985, "dur": 6.428,
    "args": {
      "External id": 522,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878055.323, "dur": 8.708,
    "args": {
      "External id": 523,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878056.188, "dur": 1.192,
    "args": {
      "External id": 524,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878057.909, "dur": 5.968,
    "args": {
      "External id": 525,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878068.286, "dur": 0.138,
    "args": {
      "External id": 526,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878109.742, "dur": 99.916,
    "args": {
      "External id": 527,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878119.282, "dur": 3.215,
    "args": {
      "External id": 528,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878121.000, "dur": 0.646,
    "args": {
      "External id": 529,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878124.127, "dur": 1.643,
    "args": {
      "External id": 530,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878125.116, "dur": 0.302,
    "args": {
      "External id": 531,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878128.145, "dur": 9.225,
    "args": {
      "External id": 532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878140.137, "dur": 10.701,
    "args": {
      "External id": 533,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878153.598, "dur": 7.355,
    "args": {
      "External id": 534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878162.452, "dur": 6.651,
    "args": {
      "External id": 535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878171.045, "dur": 7.312,
    "args": {
      "External id": 536,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878180.690, "dur": 8.455,
    "args": {
      "External id": 537,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878181.449, "dur": 0.123,
    "args": {
      "External id": 538,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878190.721, "dur": 6.987,
    "args": {
      "External id": 539,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878198.798, "dur": 7.273,
    "args": {
      "External id": 540,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878213.958, "dur": 10.307,
    "args": {
      "External id": 541,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878215.074, "dur": 1.359,
    "args": {
      "External id": 542,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878217.272, "dur": 6.843,
    "args": {
      "External id": 543,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878229.352, "dur": 0.132,
    "args": {
      "External id": 544,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878234.891, "dur": 81.582,
    "args": {
      "External id": 545,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878242.296, "dur": 2.039,
    "args": {
      "External id": 546,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878243.501, "dur": 0.387,
    "args": {
      "External id": 547,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878245.539, "dur": 1.332,
    "args": {
      "External id": 548,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878246.311, "dur": 0.259,
    "args": {
      "External id": 549,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878249.122, "dur": 6.231,
    "args": {
      "External id": 550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878257.528, "dur": 5.851,
    "args": {
      "External id": 551,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878265.443, "dur": 6.636,
    "args": {
      "External id": 552,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878273.585, "dur": 6.092,
    "args": {
      "External id": 553,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878281.275, "dur": 6.700,
    "args": {
      "External id": 554,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878290.209, "dur": 6.562,
    "args": {
      "External id": 555,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878290.812, "dur": 0.099,
    "args": {
      "External id": 556,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878298.315, "dur": 6.637,
    "args": {
      "External id": 557,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878306.033, "dur": 7.326,
    "args": {
      "External id": 558,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878319.886, "dur": 7.718,
    "args": {
      "External id": 559,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878320.759, "dur": 1.228,
    "args": {
      "External id": 560,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878322.473, "dur": 4.981,
    "args": {
      "External id": 561,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878332.453, "dur": 0.147,
    "args": {
      "External id": 562,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878337.413, "dur": 78.956,
    "args": {
      "External id": 563,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878344.307, "dur": 2.023,
    "args": {
      "External id": 564,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878345.407, "dur": 0.391,
    "args": {
      "External id": 565,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878347.607, "dur": 1.422,
    "args": {
      "External id": 566,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878348.440, "dur": 0.277,
    "args": {
      "External id": 567,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878351.019, "dur": 6.549,
    "args": {
      "External id": 568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878359.585, "dur": 6.497,
    "args": {
      "External id": 569,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878367.993, "dur": 5.918,
    "args": {
      "External id": 570,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878375.237, "dur": 5.227,
    "args": {
      "External id": 571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878381.968, "dur": 6.585,
    "args": {
      "External id": 572,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878390.745, "dur": 6.749,
    "args": {
      "External id": 573,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878391.361, "dur": 0.093,
    "args": {
      "External id": 574,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878399.032, "dur": 6.611,
    "args": {
      "External id": 575,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878406.692, "dur": 6.709,
    "args": {
      "External id": 576,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878419.648, "dur": 8.654,
    "args": {
      "External id": 577,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878420.535, "dur": 1.202,
    "args": {
      "External id": 578,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878422.256, "dur": 5.904,
    "args": {
      "External id": 579,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878432.881, "dur": 0.137,
    "args": {
      "External id": 580,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878437.874, "dur": 77.637,
    "args": {
      "External id": 581,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878444.535, "dur": 1.993,
    "args": {
      "External id": 582,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878445.627, "dur": 0.412,
    "args": {
      "External id": 583,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878447.800, "dur": 1.415,
    "args": {
      "External id": 584,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878448.649, "dur": 0.249,
    "args": {
      "External id": 585,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878451.042, "dur": 6.052,
    "args": {
      "External id": 586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878458.838, "dur": 6.519,
    "args": {
      "External id": 587,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878467.290, "dur": 6.780,
    "args": {
      "External id": 588,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878475.399, "dur": 5.947,
    "args": {
      "External id": 589,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878482.900, "dur": 6.447,
    "args": {
      "External id": 590,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878491.402, "dur": 6.500,
    "args": {
      "External id": 591,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878492.055, "dur": 0.096,
    "args": {
      "External id": 592,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878499.466, "dur": 6.177,
    "args": {
      "External id": 593,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878506.802, "dur": 5.754,
    "args": {
      "External id": 594,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878518.908, "dur": 7.681,
    "args": {
      "External id": 595,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878519.815, "dur": 1.205,
    "args": {
      "External id": 596,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878521.504, "dur": 4.922,
    "args": {
      "External id": 597,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878530.875, "dur": 0.126,
    "args": {
      "External id": 598,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878535.993, "dur": 79.220,
    "args": {
      "External id": 599,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878542.266, "dur": 1.881,
    "args": {
      "External id": 600,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878543.290, "dur": 0.418,
    "args": {
      "External id": 601,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878545.403, "dur": 1.390,
    "args": {
      "External id": 602,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878546.226, "dur": 0.242,
    "args": {
      "External id": 603,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878548.758, "dur": 6.771,
    "args": {
      "External id": 604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878557.330, "dur": 6.511,
    "args": {
      "External id": 605,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878565.693, "dur": 5.925,
    "args": {
      "External id": 606,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878573.018, "dur": 6.005,
    "args": {
      "External id": 607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878580.536, "dur": 6.383,
    "args": {
      "External id": 608,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878589.037, "dur": 6.464,
    "args": {
      "External id": 609,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878589.653, "dur": 0.100,
    "args": {
      "External id": 610,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878597.005, "dur": 6.747,
    "args": {
      "External id": 611,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878604.765, "dur": 7.464,
    "args": {
      "External id": 612,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878618.524, "dur": 8.683,
    "args": {
      "External id": 613,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878619.366, "dur": 1.221,
    "args": {
      "External id": 614,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878621.112, "dur": 5.953,
    "args": {
      "External id": 615,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878631.801, "dur": 0.148,
    "args": {
      "External id": 616,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878636.488, "dur": 78.494,
    "args": {
      "External id": 617,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878642.950, "dur": 1.859,
    "args": {
      "External id": 618,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878643.997, "dur": 0.367,
    "args": {
      "External id": 619,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878646.085, "dur": 1.371,
    "args": {
      "External id": 620,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878646.887, "dur": 0.254,
    "args": {
      "External id": 621,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878649.293, "dur": 6.581,
    "args": {
      "External id": 622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878657.617, "dur": 6.499,
    "args": {
      "External id": 623,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878665.919, "dur": 6.002,
    "args": {
      "External id": 624,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878673.268, "dur": 6.022,
    "args": {
      "External id": 625,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878680.902, "dur": 6.616,
    "args": {
      "External id": 626,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878689.677, "dur": 6.997,
    "args": {
      "External id": 627,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878690.315, "dur": 0.096,
    "args": {
      "External id": 628,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878698.231, "dur": 6.403,
    "args": {
      "External id": 629,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878705.634, "dur": 6.555,
    "args": {
      "External id": 630,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878718.220, "dur": 8.123,
    "args": {
      "External id": 631,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878719.054, "dur": 1.198,
    "args": {
      "External id": 632,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878720.786, "dur": 5.410,
    "args": {
      "External id": 633,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878730.684, "dur": 0.123,
    "args": {
      "External id": 634,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878735.374, "dur": 75.596,
    "args": {
      "External id": 635,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878741.432, "dur": 1.800,
    "args": {
      "External id": 636,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878742.434, "dur": 0.357,
    "args": {
      "External id": 637,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878744.428, "dur": 1.405,
    "args": {
      "External id": 638,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878745.255, "dur": 0.260,
    "args": {
      "External id": 639,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878748.555, "dur": 5.763,
    "args": {
      "External id": 640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878756.115, "dur": 5.917,
    "args": {
      "External id": 641,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878763.878, "dur": 5.532,
    "args": {
      "External id": 642,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878770.719, "dur": 6.281,
    "args": {
      "External id": 643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878778.558, "dur": 6.591,
    "args": {
      "External id": 644,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878787.242, "dur": 6.424,
    "args": {
      "External id": 645,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878787.851, "dur": 0.096,
    "args": {
      "External id": 646,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878795.129, "dur": 6.190,
    "args": {
      "External id": 647,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878802.309, "dur": 5.865,
    "args": {
      "External id": 648,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878814.060, "dur": 8.167,
    "args": {
      "External id": 649,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878814.871, "dur": 1.219,
    "args": {
      "External id": 650,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878816.611, "dur": 5.467,
    "args": {
      "External id": 651,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878826.798, "dur": 0.156,
    "args": {
      "External id": 652,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878831.460, "dur": 75.312,
    "args": {
      "External id": 653,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878837.408, "dur": 1.771,
    "args": {
      "External id": 654,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878838.377, "dur": 0.378,
    "args": {
      "External id": 655,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878840.368, "dur": 1.423,
    "args": {
      "External id": 656,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878841.220, "dur": 0.254,
    "args": {
      "External id": 657,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878843.606, "dur": 5.459,
    "args": {
      "External id": 658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878850.818, "dur": 6.765,
    "args": {
      "External id": 659,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878859.403, "dur": 6.260,
    "args": {
      "External id": 660,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878866.984, "dur": 5.671,
    "args": {
      "External id": 661,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878874.203, "dur": 6.407,
    "args": {
      "External id": 662,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878882.716, "dur": 6.319,
    "args": {
      "External id": 663,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878883.322, "dur": 0.089,
    "args": {
      "External id": 664,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878890.574, "dur": 6.734,
    "args": {
      "External id": 665,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878898.318, "dur": 5.592,
    "args": {
      "External id": 666,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740878909.969, "dur": 8.290,
    "args": {
      "External id": 667,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878910.897, "dur": 1.260,
    "args": {
      "External id": 668,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740878912.667, "dur": 5.406,
    "args": {
      "External id": 669,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878922.660, "dur": 0.148,
    "args": {
      "External id": 670,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878927.571, "dur": 76.993,
    "args": {
      "External id": 671,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878933.468, "dur": 1.857,
    "args": {
      "External id": 672,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878934.494, "dur": 0.394,
    "args": {
      "External id": 673,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740878936.597, "dur": 1.384,
    "args": {
      "External id": 674,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740878937.415, "dur": 0.254,
    "args": {
      "External id": 675,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740878939.801, "dur": 6.885,
    "args": {
      "External id": 676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878948.549, "dur": 6.232,
    "args": {
      "External id": 677,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878957.652, "dur": 5.982,
    "args": {
      "External id": 678,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740878964.926, "dur": 5.557,
    "args": {
      "External id": 679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740878971.923, "dur": 6.019,
    "args": {
      "External id": 680,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740878980.005, "dur": 6.652,
    "args": {
      "External id": 681,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740878980.700, "dur": 0.098,
    "args": {
      "External id": 682,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740878988.127, "dur": 6.468,
    "args": {
      "External id": 683,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740878995.555, "dur": 6.320,
    "args": {
      "External id": 684,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879007.842, "dur": 8.416,
    "args": {
      "External id": 685,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879008.672, "dur": 1.188,
    "args": {
      "External id": 686,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879010.359, "dur": 5.749,
    "args": {
      "External id": 687,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879020.451, "dur": 0.127,
    "args": {
      "External id": 688,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879025.472, "dur": 77.849,
    "args": {
      "External id": 689,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879031.448, "dur": 1.939,
    "args": {
      "External id": 690,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879032.560, "dur": 0.377,
    "args": {
      "External id": 691,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879034.545, "dur": 1.389,
    "args": {
      "External id": 692,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879035.366, "dur": 0.259,
    "args": {
      "External id": 693,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879037.773, "dur": 7.068,
    "args": {
      "External id": 694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879046.650, "dur": 5.554,
    "args": {
      "External id": 695,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879054.067, "dur": 6.071,
    "args": {
      "External id": 696,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879061.487, "dur": 6.130,
    "args": {
      "External id": 697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879069.095, "dur": 7.045,
    "args": {
      "External id": 698,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879078.119, "dur": 7.563,
    "args": {
      "External id": 699,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879079.749, "dur": 0.101,
    "args": {
      "External id": 700,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879087.111, "dur": 6.146,
    "args": {
      "External id": 701,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879094.181, "dur": 6.286,
    "args": {
      "External id": 702,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879106.601, "dur": 8.749,
    "args": {
      "External id": 703,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879107.390, "dur": 1.215,
    "args": {
      "External id": 704,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879109.039, "dur": 6.134,
    "args": {
      "External id": 705,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879119.581, "dur": 0.129,
    "args": {
      "External id": 706,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879126.477, "dur": 76.699,
    "args": {
      "External id": 707,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879132.482, "dur": 1.844,
    "args": {
      "External id": 708,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879133.530, "dur": 0.358,
    "args": {
      "External id": 709,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879135.542, "dur": 1.422,
    "args": {
      "External id": 710,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879136.402, "dur": 0.268,
    "args": {
      "External id": 711,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879138.711, "dur": 6.019,
    "args": {
      "External id": 712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879146.449, "dur": 7.416,
    "args": {
      "External id": 713,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879155.683, "dur": 5.990,
    "args": {
      "External id": 714,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879162.976, "dur": 5.777,
    "args": {
      "External id": 715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879170.258, "dur": 6.384,
    "args": {
      "External id": 716,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879178.810, "dur": 6.860,
    "args": {
      "External id": 717,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879179.612, "dur": 0.088,
    "args": {
      "External id": 718,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879187.040, "dur": 6.278,
    "args": {
      "External id": 719,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879194.266, "dur": 6.122,
    "args": {
      "External id": 720,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879206.430, "dur": 8.330,
    "args": {
      "External id": 721,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879207.312, "dur": 1.184,
    "args": {
      "External id": 722,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879208.948, "dur": 5.656,
    "args": {
      "External id": 723,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879218.991, "dur": 0.125,
    "args": {
      "External id": 724,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879223.535, "dur": 95.516,
    "args": {
      "External id": 725,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879229.434, "dur": 1.833,
    "args": {
      "External id": 726,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879230.476, "dur": 0.359,
    "args": {
      "External id": 727,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879232.444, "dur": 1.421,
    "args": {
      "External id": 728,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879233.274, "dur": 0.275,
    "args": {
      "External id": 729,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879235.579, "dur": 5.897,
    "args": {
      "External id": 730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879243.114, "dur": 6.923,
    "args": {
      "External id": 731,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879251.839, "dur": 6.453,
    "args": {
      "External id": 732,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879259.595, "dur": 18.339,
    "args": {
      "External id": 733,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879279.592, "dur": 6.336,
    "args": {
      "External id": 734,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879287.995, "dur": 7.558,
    "args": {
      "External id": 735,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879288.805, "dur": 0.086,
    "args": {
      "External id": 736,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879297.027, "dur": 6.212,
    "args": {
      "External id": 737,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879308.969, "dur": 7.174,
    "args": {
      "External id": 738,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879322.256, "dur": 8.313,
    "args": {
      "External id": 739,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879323.151, "dur": 1.235,
    "args": {
      "External id": 740,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879324.886, "dur": 5.532,
    "args": {
      "External id": 741,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879334.828, "dur": 0.131,
    "args": {
      "External id": 742,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879339.517, "dur": 81.909,
    "args": {
      "External id": 743,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879345.528, "dur": 1.972,
    "args": {
      "External id": 744,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879346.649, "dur": 0.351,
    "args": {
      "External id": 745,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879348.713, "dur": 1.499,
    "args": {
      "External id": 746,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879349.648, "dur": 0.261,
    "args": {
      "External id": 747,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879351.959, "dur": 7.343,
    "args": {
      "External id": 748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879361.457, "dur": 6.752,
    "args": {
      "External id": 749,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879369.952, "dur": 6.629,
    "args": {
      "External id": 750,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879377.963, "dur": 5.723,
    "args": {
      "External id": 751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879385.243, "dur": 7.626,
    "args": {
      "External id": 752,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879395.027, "dur": 6.839,
    "args": {
      "External id": 753,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879395.725, "dur": 0.096,
    "args": {
      "External id": 754,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879403.289, "dur": 7.070,
    "args": {
      "External id": 755,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879411.329, "dur": 7.252,
    "args": {
      "External id": 756,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879424.656, "dur": 9.533,
    "args": {
      "External id": 757,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879425.443, "dur": 1.215,
    "args": {
      "External id": 758,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879427.144, "dur": 6.895,
    "args": {
      "External id": 759,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879438.437, "dur": 0.132,
    "args": {
      "External id": 760,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879443.166, "dur": 76.532,
    "args": {
      "External id": 761,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879448.867, "dur": 1.860,
    "args": {
      "External id": 762,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879449.903, "dur": 0.381,
    "args": {
      "External id": 763,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879451.852, "dur": 1.401,
    "args": {
      "External id": 764,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879452.670, "dur": 0.264,
    "args": {
      "External id": 765,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879454.977, "dur": 5.623,
    "args": {
      "External id": 766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879462.368, "dur": 7.615,
    "args": {
      "External id": 767,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879471.887, "dur": 6.553,
    "args": {
      "External id": 768,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879479.769, "dur": 5.843,
    "args": {
      "External id": 769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879487.125, "dur": 6.730,
    "args": {
      "External id": 770,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879495.879, "dur": 6.722,
    "args": {
      "External id": 771,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879496.450, "dur": 0.095,
    "args": {
      "External id": 772,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879503.988, "dur": 6.363,
    "args": {
      "External id": 773,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879511.269, "dur": 5.583,
    "args": {
      "External id": 774,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879522.777, "dur": 7.615,
    "args": {
      "External id": 775,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879523.612, "dur": 1.188,
    "args": {
      "External id": 776,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879525.295, "dur": 4.953,
    "args": {
      "External id": 777,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879535.200, "dur": 0.103,
    "args": {
      "External id": 778,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879539.618, "dur": 77.060,
    "args": {
      "External id": 779,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879545.569, "dur": 1.978,
    "args": {
      "External id": 780,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879546.745, "dur": 0.360,
    "args": {
      "External id": 781,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879548.691, "dur": 1.476,
    "args": {
      "External id": 782,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879549.608, "dur": 0.261,
    "args": {
      "External id": 783,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879551.873, "dur": 6.640,
    "args": {
      "External id": 784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879560.321, "dur": 6.657,
    "args": {
      "External id": 785,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879568.917, "dur": 5.899,
    "args": {
      "External id": 786,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879576.135, "dur": 6.129,
    "args": {
      "External id": 787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879583.789, "dur": 6.335,
    "args": {
      "External id": 788,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879592.349, "dur": 5.980,
    "args": {
      "External id": 789,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879592.963, "dur": 0.080,
    "args": {
      "External id": 790,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879599.843, "dur": 6.338,
    "args": {
      "External id": 791,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879607.235, "dur": 6.500,
    "args": {
      "External id": 792,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879619.856, "dur": 8.197,
    "args": {
      "External id": 793,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879620.737, "dur": 1.254,
    "args": {
      "External id": 794,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879622.482, "dur": 5.428,
    "args": {
      "External id": 795,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879632.380, "dur": 0.138,
    "args": {
      "External id": 796,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879636.822, "dur": 76.810,
    "args": {
      "External id": 797,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879642.663, "dur": 1.862,
    "args": {
      "External id": 798,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879643.704, "dur": 0.383,
    "args": {
      "External id": 799,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879645.669, "dur": 1.427,
    "args": {
      "External id": 800,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879646.517, "dur": 0.268,
    "args": {
      "External id": 801,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879648.928, "dur": 6.745,
    "args": {
      "External id": 802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879657.516, "dur": 6.125,
    "args": {
      "External id": 803,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879665.381, "dur": 5.961,
    "args": {
      "External id": 804,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879672.633, "dur": 5.393,
    "args": {
      "External id": 805,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879679.523, "dur": 7.082,
    "args": {
      "External id": 806,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879688.812, "dur": 7.187,
    "args": {
      "External id": 807,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879689.414, "dur": 0.085,
    "args": {
      "External id": 808,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879697.445, "dur": 6.053,
    "args": {
      "External id": 809,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879704.507, "dur": 6.317,
    "args": {
      "External id": 810,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879716.803, "dur": 8.367,
    "args": {
      "External id": 811,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879717.692, "dur": 1.175,
    "args": {
      "External id": 812,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879719.349, "dur": 5.684,
    "args": {
      "External id": 813,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879729.472, "dur": 0.143,
    "args": {
      "External id": 814,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879733.789, "dur": 77.337,
    "args": {
      "External id": 815,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879739.550, "dur": 1.888,
    "args": {
      "External id": 816,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879740.617, "dur": 0.377,
    "args": {
      "External id": 817,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879742.584, "dur": 1.393,
    "args": {
      "External id": 818,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879743.412, "dur": 0.261,
    "args": {
      "External id": 819,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879745.795, "dur": 6.415,
    "args": {
      "External id": 820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879754.011, "dur": 7.311,
    "args": {
      "External id": 821,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879763.267, "dur": 5.564,
    "args": {
      "External id": 822,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879770.096, "dur": 5.980,
    "args": {
      "External id": 823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879777.573, "dur": 6.789,
    "args": {
      "External id": 824,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879786.427, "dur": 6.831,
    "args": {
      "External id": 825,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879787.156, "dur": 0.096,
    "args": {
      "External id": 826,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879794.658, "dur": 6.207,
    "args": {
      "External id": 827,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879801.814, "dur": 6.628,
    "args": {
      "External id": 828,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879814.230, "dur": 8.408,
    "args": {
      "External id": 829,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879815.116, "dur": 1.186,
    "args": {
      "External id": 830,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879816.775, "dur": 5.720,
    "args": {
      "External id": 831,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879826.763, "dur": 0.139,
    "args": {
      "External id": 832,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879831.210, "dur": 73.056,
    "args": {
      "External id": 833,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879837.065, "dur": 1.881,
    "args": {
      "External id": 834,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879838.124, "dur": 0.383,
    "args": {
      "External id": 835,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879840.105, "dur": 1.450,
    "args": {
      "External id": 836,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879840.980, "dur": 0.263,
    "args": {
      "External id": 837,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879843.228, "dur": 5.661,
    "args": {
      "External id": 838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879850.636, "dur": 5.642,
    "args": {
      "External id": 839,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879857.983, "dur": 5.353,
    "args": {
      "External id": 840,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879864.539, "dur": 6.374,
    "args": {
      "External id": 841,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879872.508, "dur": 6.305,
    "args": {
      "External id": 842,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879880.838, "dur": 6.398,
    "args": {
      "External id": 843,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879881.415, "dur": 0.100,
    "args": {
      "External id": 844,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879888.637, "dur": 6.264,
    "args": {
      "External id": 845,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879895.851, "dur": 5.625,
    "args": {
      "External id": 846,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740879907.284, "dur": 8.012,
    "args": {
      "External id": 847,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879908.153, "dur": 1.154,
    "args": {
      "External id": 848,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740879909.811, "dur": 5.339,
    "args": {
      "External id": 849,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879919.398, "dur": 0.145,
    "args": {
      "External id": 850,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879923.824, "dur": 75.063,
    "args": {
      "External id": 851,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879929.549, "dur": 1.865,
    "args": {
      "External id": 852,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879930.616, "dur": 0.387,
    "args": {
      "External id": 853,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740879932.655, "dur": 1.413,
    "args": {
      "External id": 854,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740879933.485, "dur": 0.272,
    "args": {
      "External id": 855,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740879935.795, "dur": 6.033,
    "args": {
      "External id": 856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879943.615, "dur": 6.916,
    "args": {
      "External id": 857,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879952.386, "dur": 6.160,
    "args": {
      "External id": 858,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740879959.806, "dur": 5.646,
    "args": {
      "External id": 859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740879967.025, "dur": 6.540,
    "args": {
      "External id": 860,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740879975.592, "dur": 6.209,
    "args": {
      "External id": 861,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740879976.224, "dur": 0.095,
    "args": {
      "External id": 862,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740879983.209, "dur": 5.715,
    "args": {
      "External id": 863,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740879989.874, "dur": 6.285,
    "args": {
      "External id": 864,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740880001.967, "dur": 8.171,
    "args": {
      "External id": 865,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880002.797, "dur": 1.140,
    "args": {
      "External id": 866,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740880004.421, "dur": 5.561,
    "args": {
      "External id": 867,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740880014.347, "dur": 0.127,
    "args": {
      "External id": 868,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740880018.646, "dur": 74.952,
    "args": {
      "External id": 869,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740880024.396, "dur": 2.084,
    "args": {
      "External id": 870,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880025.702, "dur": 0.364,
    "args": {
      "External id": 871,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740880027.659, "dur": 1.445,
    "args": {
      "External id": 872,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880028.536, "dur": 0.268,
    "args": {
      "External id": 873,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740880030.967, "dur": 6.131,
    "args": {
      "External id": 874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880038.910, "dur": 6.375,
    "args": {
      "External id": 875,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880047.145, "dur": 5.753,
    "args": {
      "External id": 876,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740880054.186, "dur": 5.507,
    "args": {
      "External id": 877,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740880061.228, "dur": 6.222,
    "args": {
      "External id": 878,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740880069.439, "dur": 6.039,
    "args": {
      "External id": 879,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740880070.026, "dur": 0.091,
    "args": {
      "External id": 880,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740880076.953, "dur": 6.653,
    "args": {
      "External id": 881,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880084.648, "dur": 6.224,
    "args": {
      "External id": 882,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740880096.732, "dur": 8.381,
    "args": {
      "External id": 883,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880097.638, "dur": 1.171,
    "args": {
      "External id": 884,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740880099.318, "dur": 5.640,
    "args": {
      "External id": 885,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740880109.287, "dur": 0.153,
    "args": {
      "External id": 886,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740880114.159, "dur": 76.874,
    "args": {
      "External id": 887,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740880119.758, "dur": 1.922,
    "args": {
      "External id": 888,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880120.885, "dur": 0.361,
    "args": {
      "External id": 889,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740880122.872, "dur": 1.438,
    "args": {
      "External id": 890,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740880123.730, "dur": 0.265,
    "args": {
      "External id": 891,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740880126.218, "dur": 6.185,
    "args": {
      "External id": 892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880134.216, "dur": 6.040,
    "args": {
      "External id": 893,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880142.042, "dur": 5.920,
    "args": {
      "External id": 894,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740880149.283, "dur": 6.251,
    "args": {
      "External id": 895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740880157.110, "dur": 6.854,
    "args": {
      "External id": 896,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740880166.001, "dur": 6.551,
    "args": {
      "External id": 897,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740880166.604, "dur": 0.109,
    "args": {
      "External id": 898,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740880173.979, "dur": 6.150,
    "args": {
      "External id": 899,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740880182.121, "dur": 6.168,
    "args": {
      "External id": 900,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 49625, "tid": 49625,
    "ts": 7253740901446.242, "dur": 183.775,
    "args": {
      "External id": 901,"Record function id": 0, "Concrete Inputs": ["[67108864]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740901643.780, "dur": 28.771,
    "args": {
      "External id": 902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740901651.415, "dur": 20.388,
    "args": {
      "External id": 903,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740901682.452, "dur": 23.491,
    "args": {
      "External id": 904,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740901685.996, "dur": 3.475,
    "args": {
      "External id": 905,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740901691.056, "dur": 14.618,
    "args": {
      "External id": 906,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740901783.766, "dur": 0.457,
    "args": {
      "External id": 907,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740901800.057, "dur": 172.910,
    "args": {
      "External id": 908,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740901821.391, "dur": 12.526,
    "args": {
      "External id": 909,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740901828.574, "dur": 2.692,
    "args": {
      "External id": 910,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740901835.840, "dur": 2.329,
    "args": {
      "External id": 911,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740901837.060, "dur": 0.682,
    "args": {
      "External id": 912,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740901843.994, "dur": 17.036,
    "args": {
      "External id": 913,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740901868.593, "dur": 18.198,
    "args": {
      "External id": 914,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740901891.214, "dur": 9.848,
    "args": {
      "External id": 915,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740901903.812, "dur": 10.093,
    "args": {
      "External id": 916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740901917.858, "dur": 11.035,
    "args": {
      "External id": 917,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740901933.469, "dur": 12.946,
    "args": {
      "External id": 918,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740901936.145, "dur": 0.192,
    "args": {
      "External id": 919,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740901949.776, "dur": 8.681,
    "args": {
      "External id": 920,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740901959.951, "dur": 6.915,
    "args": {
      "External id": 921,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902007.938, "dur": 9.098,
    "args": {
      "External id": 922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902009.411, "dur": 7.208,
    "args": {
      "External id": 923,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902022.516, "dur": 14.182,
    "args": {
      "External id": 924,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902024.047, "dur": 2.998,
    "args": {
      "External id": 925,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902027.882, "dur": 8.594,
    "args": {
      "External id": 926,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902087.765, "dur": 0.467,
    "args": {
      "External id": 927,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902097.729, "dur": 109.093,
    "args": {
      "External id": 928,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902108.744, "dur": 4.058,
    "args": {
      "External id": 929,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902110.980, "dur": 0.895,
    "args": {
      "External id": 930,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902114.804, "dur": 1.717,
    "args": {
      "External id": 931,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902115.748, "dur": 0.424,
    "args": {
      "External id": 932,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740902119.653, "dur": 11.327,
    "args": {
      "External id": 933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902135.221, "dur": 9.391,
    "args": {
      "External id": 934,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902148.133, "dur": 7.611,
    "args": {
      "External id": 935,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740902157.373, "dur": 7.060,
    "args": {
      "External id": 936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740902166.512, "dur": 7.704,
    "args": {
      "External id": 937,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740902176.823, "dur": 8.080,
    "args": {
      "External id": 938,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902177.953, "dur": 0.244,
    "args": {
      "External id": 939,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740902186.842, "dur": 7.012,
    "args": {
      "External id": 940,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902195.433, "dur": 6.972,
    "args": {
      "External id": 941,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902234.315, "dur": 7.792,
    "args": {
      "External id": 942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902235.645, "dur": 6.114,
    "args": {
      "External id": 943,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902246.425, "dur": 10.697,
    "args": {
      "External id": 944,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902247.595, "dur": 2.007,
    "args": {
      "External id": 945,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902250.158, "dur": 6.791,
    "args": {
      "External id": 946,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902282.505, "dur": 0.230,
    "args": {
      "External id": 947,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902292.348, "dur": 100.941,
    "args": {
      "External id": 948,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902301.402, "dur": 3.052,
    "args": {
      "External id": 949,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902302.748, "dur": 1.126,
    "args": {
      "External id": 950,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902305.790, "dur": 2.515,
    "args": {
      "External id": 951,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902307.644, "dur": 0.303,
    "args": {
      "External id": 952,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740902311.010, "dur": 8.209,
    "args": {
      "External id": 953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902323.795, "dur": 10.110,
    "args": {
      "External id": 954,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902336.585, "dur": 7.446,
    "args": {
      "External id": 955,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740902345.434, "dur": 6.505,
    "args": {
      "External id": 956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740902353.781, "dur": 7.260,
    "args": {
      "External id": 957,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740902363.575, "dur": 8.598,
    "args": {
      "External id": 958,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902365.265, "dur": 0.102,
    "args": {
      "External id": 959,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740902374.022, "dur": 7.346,
    "args": {
      "External id": 960,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902382.713, "dur": 6.944,
    "args": {
      "External id": 961,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902416.359, "dur": 6.437,
    "args": {
      "External id": 962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902417.220, "dur": 5.284,
    "args": {
      "External id": 963,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902426.662, "dur": 9.428,
    "args": {
      "External id": 964,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902427.595, "dur": 2.072,
    "args": {
      "External id": 965,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902430.183, "dur": 5.765,
    "args": {
      "External id": 966,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902457.537, "dur": 0.172,
    "args": {
      "External id": 967,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902463.535, "dur": 84.826,
    "args": {
      "External id": 968,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902472.079, "dur": 2.301,
    "args": {
      "External id": 969,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902473.396, "dur": 0.451,
    "args": {
      "External id": 970,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902475.713, "dur": 1.617,
    "args": {
      "External id": 971,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902476.657, "dur": 0.369,
    "args": {
      "External id": 972,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740902479.605, "dur": 7.400,
    "args": {
      "External id": 973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902489.174, "dur": 7.310,
    "args": {
      "External id": 974,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902498.814, "dur": 5.870,
    "args": {
      "External id": 975,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740902506.053, "dur": 5.959,
    "args": {
      "External id": 976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740902513.838, "dur": 6.451,
    "args": {
      "External id": 977,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740902522.651, "dur": 7.216,
    "args": {
      "External id": 978,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902524.037, "dur": 0.120,
    "args": {
      "External id": 979,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740902531.514, "dur": 6.461,
    "args": {
      "External id": 980,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902538.985, "dur": 5.933,
    "args": {
      "External id": 981,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902569.145, "dur": 6.148,
    "args": {
      "External id": 982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902569.995, "dur": 4.988,
    "args": {
      "External id": 983,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902579.182, "dur": 9.014,
    "args": {
      "External id": 984,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902580.075, "dur": 1.938,
    "args": {
      "External id": 985,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902582.552, "dur": 5.497,
    "args": {
      "External id": 986,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902610.230, "dur": 0.168,
    "args": {
      "External id": 987,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902616.066, "dur": 100.412,
    "args": {
      "External id": 988,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902623.963, "dur": 2.856,
    "args": {
      "External id": 989,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902625.876, "dur": 0.427,
    "args": {
      "External id": 990,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902628.146, "dur": 1.679,
    "args": {
      "External id": 991,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902629.023, "dur": 0.445,
    "args": {
      "External id": 992,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740902632.096, "dur": 6.954,
    "args": {
      "External id": 993,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902641.096, "dur": 7.419,
    "args": {
      "External id": 994,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902650.659, "dur": 5.975,
    "args": {
      "External id": 995,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740902657.980, "dur": 5.348,
    "args": {
      "External id": 996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740902665.195, "dur": 7.876,
    "args": {
      "External id": 997,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740902675.501, "dur": 11.089,
    "args": {
      "External id": 998,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902677.490, "dur": 0.173,
    "args": {
      "External id": 999,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740902688.311, "dur": 7.633,
    "args": {
      "External id": 1000,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902697.040, "dur": 14.376,
    "args": {
      "External id": 1001,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902741.648, "dur": 7.604,
    "args": {
      "External id": 1002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902742.983, "dur": 5.974,
    "args": {
      "External id": 1003,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902753.436, "dur": 11.900,
    "args": {
      "External id": 1004,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902755.430, "dur": 2.608,
    "args": {
      "External id": 1005,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902758.828, "dur": 6.286,
    "args": {
      "External id": 1006,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902787.768, "dur": 0.208,
    "args": {
      "External id": 1007,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902796.097, "dur": 104.418,
    "args": {
      "External id": 1008,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902805.068, "dur": 2.652,
    "args": {
      "External id": 1009,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902806.478, "dur": 0.558,
    "args": {
      "External id": 1010,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740902809.093, "dur": 1.870,
    "args": {
      "External id": 1011,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902810.182, "dur": 0.383,
    "args": {
      "External id": 1012,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740902813.346, "dur": 17.857,
    "args": {
      "External id": 1013,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902834.353, "dur": 8.176,
    "args": {
      "External id": 1014,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902845.017, "dur": 7.162,
    "args": {
      "External id": 1015,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740902853.677, "dur": 6.857,
    "args": {
      "External id": 1016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740902862.351, "dur": 7.997,
    "args": {
      "External id": 1017,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740902872.789, "dur": 8.273,
    "args": {
      "External id": 1018,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902873.618, "dur": 0.215,
    "args": {
      "External id": 1019,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740902882.811, "dur": 6.875,
    "args": {
      "External id": 1020,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740902890.700, "dur": 6.379,
    "args": {
      "External id": 1021,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740902925.648, "dur": 6.507,
    "args": {
      "External id": 1022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740902926.615, "dur": 5.254,
    "args": {
      "External id": 1023,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740902936.143, "dur": 33.968,
    "args": {
      "External id": 1024,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740902959.303, "dur": 3.682,
    "args": {
      "External id": 1025,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740902963.491, "dur": 6.495,
    "args": {
      "External id": 1026,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740902992.516, "dur": 0.280,
    "args": {
      "External id": 1027,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903001.722, "dur": 94.631,
    "args": {
      "External id": 1028,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903012.069, "dur": 3.542,
    "args": {
      "External id": 1029,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903014.492, "dur": 0.588,
    "args": {
      "External id": 1030,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903017.077, "dur": 1.816,
    "args": {
      "External id": 1031,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903018.191, "dur": 0.392,
    "args": {
      "External id": 1032,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903021.404, "dur": 7.614,
    "args": {
      "External id": 1033,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903031.800, "dur": 8.226,
    "args": {
      "External id": 1034,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903042.580, "dur": 6.733,
    "args": {
      "External id": 1035,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903050.825, "dur": 6.369,
    "args": {
      "External id": 1036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903059.045, "dur": 7.399,
    "args": {
      "External id": 1037,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903068.762, "dur": 7.936,
    "args": {
      "External id": 1038,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903070.121, "dur": 0.127,
    "args": {
      "External id": 1039,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903078.297, "dur": 6.912,
    "args": {
      "External id": 1040,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903086.196, "dur": 6.338,
    "args": {
      "External id": 1041,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903119.477, "dur": 7.106,
    "args": {
      "External id": 1042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903120.418, "dur": 5.891,
    "args": {
      "External id": 1043,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903130.445, "dur": 9.614,
    "args": {
      "External id": 1044,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903132.134, "dur": 1.851,
    "args": {
      "External id": 1045,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903134.462, "dur": 5.455,
    "args": {
      "External id": 1046,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903161.753, "dur": 0.135,
    "args": {
      "External id": 1047,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903168.262, "dur": 88.901,
    "args": {
      "External id": 1048,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903178.132, "dur": 2.374,
    "args": {
      "External id": 1049,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903179.413, "dur": 0.580,
    "args": {
      "External id": 1050,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903181.893, "dur": 1.730,
    "args": {
      "External id": 1051,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903182.808, "dur": 0.426,
    "args": {
      "External id": 1052,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903186.009, "dur": 6.582,
    "args": {
      "External id": 1053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903195.080, "dur": 8.233,
    "args": {
      "External id": 1054,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903205.803, "dur": 7.120,
    "args": {
      "External id": 1055,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903214.390, "dur": 6.118,
    "args": {
      "External id": 1056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903222.235, "dur": 6.695,
    "args": {
      "External id": 1057,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903231.045, "dur": 7.149,
    "args": {
      "External id": 1058,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903232.451, "dur": 0.087,
    "args": {
      "External id": 1059,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903240.032, "dur": 6.936,
    "args": {
      "External id": 1060,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903248.082, "dur": 5.473,
    "args": {
      "External id": 1061,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903282.190, "dur": 7.482,
    "args": {
      "External id": 1062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903283.131, "dur": 6.305,
    "args": {
      "External id": 1063,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903293.435, "dur": 9.538,
    "args": {
      "External id": 1064,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903294.539, "dur": 1.846,
    "args": {
      "External id": 1065,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903296.925, "dur": 5.904,
    "args": {
      "External id": 1066,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903326.947, "dur": 0.194,
    "args": {
      "External id": 1067,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903332.965, "dur": 90.166,
    "args": {
      "External id": 1068,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903341.811, "dur": 2.784,
    "args": {
      "External id": 1069,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903343.732, "dur": 0.417,
    "args": {
      "External id": 1070,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903345.948, "dur": 1.722,
    "args": {
      "External id": 1071,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903346.867, "dur": 0.404,
    "args": {
      "External id": 1072,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903350.156, "dur": 7.523,
    "args": {
      "External id": 1073,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903359.953, "dur": 8.049,
    "args": {
      "External id": 1074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903370.374, "dur": 7.567,
    "args": {
      "External id": 1075,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903379.370, "dur": 6.447,
    "args": {
      "External id": 1076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903387.549, "dur": 7.070,
    "args": {
      "External id": 1077,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903396.918, "dur": 7.852,
    "args": {
      "External id": 1078,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903398.206, "dur": 0.103,
    "args": {
      "External id": 1079,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903406.407, "dur": 5.930,
    "args": {
      "External id": 1080,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903413.414, "dur": 6.366,
    "args": {
      "External id": 1081,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903446.383, "dur": 7.151,
    "args": {
      "External id": 1082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903447.439, "dur": 5.862,
    "args": {
      "External id": 1083,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903457.020, "dur": 9.403,
    "args": {
      "External id": 1084,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903458.584, "dur": 1.498,
    "args": {
      "External id": 1085,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903460.613, "dur": 5.668,
    "args": {
      "External id": 1086,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903488.326, "dur": 0.185,
    "args": {
      "External id": 1087,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903494.720, "dur": 87.826,
    "args": {
      "External id": 1088,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903503.495, "dur": 2.789,
    "args": {
      "External id": 1089,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903505.267, "dur": 0.505,
    "args": {
      "External id": 1090,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903507.558, "dur": 2.989,
    "args": {
      "External id": 1091,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903508.937, "dur": 1.142,
    "args": {
      "External id": 1092,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903512.765, "dur": 6.699,
    "args": {
      "External id": 1093,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903521.631, "dur": 6.179,
    "args": {
      "External id": 1094,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903529.983, "dur": 6.395,
    "args": {
      "External id": 1095,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903537.870, "dur": 7.814,
    "args": {
      "External id": 1096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903547.430, "dur": 7.791,
    "args": {
      "External id": 1097,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903557.519, "dur": 6.839,
    "args": {
      "External id": 1098,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903558.216, "dur": 0.098,
    "args": {
      "External id": 1099,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903565.958, "dur": 6.330,
    "args": {
      "External id": 1100,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903573.453, "dur": 5.822,
    "args": {
      "External id": 1101,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903605.640, "dur": 6.235,
    "args": {
      "External id": 1102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903606.559, "dur": 5.084,
    "args": {
      "External id": 1103,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903615.510, "dur": 9.114,
    "args": {
      "External id": 1104,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903616.471, "dur": 1.470,
    "args": {
      "External id": 1105,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903618.464, "dur": 6.025,
    "args": {
      "External id": 1106,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903646.588, "dur": 0.164,
    "args": {
      "External id": 1107,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903652.416, "dur": 81.653,
    "args": {
      "External id": 1108,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903661.091, "dur": 2.246,
    "args": {
      "External id": 1109,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903662.308, "dur": 0.475,
    "args": {
      "External id": 1110,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903664.780, "dur": 1.450,
    "args": {
      "External id": 1111,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903665.601, "dur": 0.308,
    "args": {
      "External id": 1112,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903668.379, "dur": 6.455,
    "args": {
      "External id": 1113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903677.072, "dur": 6.936,
    "args": {
      "External id": 1114,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903686.173, "dur": 6.145,
    "args": {
      "External id": 1115,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903693.698, "dur": 5.489,
    "args": {
      "External id": 1116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903700.822, "dur": 6.389,
    "args": {
      "External id": 1117,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903709.413, "dur": 6.132,
    "args": {
      "External id": 1118,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903710.012, "dur": 0.097,
    "args": {
      "External id": 1119,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903717.143, "dur": 6.706,
    "args": {
      "External id": 1120,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903724.870, "dur": 5.912,
    "args": {
      "External id": 1121,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903756.230, "dur": 6.853,
    "args": {
      "External id": 1122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903757.321, "dur": 5.508,
    "args": {
      "External id": 1123,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903766.797, "dur": 8.677,
    "args": {
      "External id": 1124,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903767.674, "dur": 1.587,
    "args": {
      "External id": 1125,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903769.767, "dur": 5.582,
    "args": {
      "External id": 1126,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903797.406, "dur": 0.147,
    "args": {
      "External id": 1127,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903803.388, "dur": 81.795,
    "args": {
      "External id": 1128,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903810.590, "dur": 2.420,
    "args": {
      "External id": 1129,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903811.963, "dur": 0.574,
    "args": {
      "External id": 1130,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903814.335, "dur": 1.511,
    "args": {
      "External id": 1131,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903815.131, "dur": 0.408,
    "args": {
      "External id": 1132,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903818.017, "dur": 6.265,
    "args": {
      "External id": 1133,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903826.612, "dur": 6.379,
    "args": {
      "External id": 1134,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903835.066, "dur": 6.342,
    "args": {
      "External id": 1135,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740903842.771, "dur": 5.626,
    "args": {
      "External id": 1136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740903850.028, "dur": 6.284,
    "args": {
      "External id": 1137,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740903858.599, "dur": 6.930,
    "args": {
      "External id": 1138,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903859.365, "dur": 0.089,
    "args": {
      "External id": 1139,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740903868.541, "dur": 6.323,
    "args": {
      "External id": 1140,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903875.962, "dur": 6.040,
    "args": {
      "External id": 1141,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740903906.486, "dur": 6.217,
    "args": {
      "External id": 1142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740903907.438, "dur": 4.985,
    "args": {
      "External id": 1143,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740903916.270, "dur": 8.867,
    "args": {
      "External id": 1144,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903917.441, "dur": 1.467,
    "args": {
      "External id": 1145,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740903919.499, "dur": 5.496,
    "args": {
      "External id": 1146,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740903946.273, "dur": 0.164,
    "args": {
      "External id": 1147,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903951.933, "dur": 96.715,
    "args": {
      "External id": 1148,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903959.059, "dur": 2.344,
    "args": {
      "External id": 1149,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903960.423, "dur": 0.466,
    "args": {
      "External id": 1150,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740903962.710, "dur": 1.436,
    "args": {
      "External id": 1151,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740903963.549, "dur": 0.298,
    "args": {
      "External id": 1152,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740903966.219, "dur": 22.469,
    "args": {
      "External id": 1153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903991.105, "dur": 6.601,
    "args": {
      "External id": 1154,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740903999.969, "dur": 5.913,
    "args": {
      "External id": 1155,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904007.312, "dur": 6.260,
    "args": {
      "External id": 1156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904015.263, "dur": 6.227,
    "args": {
      "External id": 1157,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904023.829, "dur": 6.484,
    "args": {
      "External id": 1158,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904024.516, "dur": 0.103,
    "args": {
      "External id": 1159,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904032.052, "dur": 5.967,
    "args": {
      "External id": 1160,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904039.156, "dur": 6.012,
    "args": {
      "External id": 1161,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904069.152, "dur": 7.958,
    "args": {
      "External id": 1162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904070.177, "dur": 6.681,
    "args": {
      "External id": 1163,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904080.779, "dur": 9.667,
    "args": {
      "External id": 1164,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904081.883, "dur": 2.357,
    "args": {
      "External id": 1165,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904084.859, "dur": 5.378,
    "args": {
      "External id": 1166,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904111.922, "dur": 0.144,
    "args": {
      "External id": 1167,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904118.359, "dur": 83.660,
    "args": {
      "External id": 1168,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904125.737, "dur": 2.591,
    "args": {
      "External id": 1169,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904127.319, "dur": 0.452,
    "args": {
      "External id": 1170,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904129.721, "dur": 1.569,
    "args": {
      "External id": 1171,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904130.576, "dur": 0.395,
    "args": {
      "External id": 1172,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904133.449, "dur": 6.669,
    "args": {
      "External id": 1173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904142.372, "dur": 7.149,
    "args": {
      "External id": 1174,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904151.675, "dur": 6.713,
    "args": {
      "External id": 1175,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904159.800, "dur": 6.366,
    "args": {
      "External id": 1176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904167.800, "dur": 6.691,
    "args": {
      "External id": 1177,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904176.936, "dur": 7.487,
    "args": {
      "External id": 1178,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904178.638, "dur": 0.095,
    "args": {
      "External id": 1179,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904186.030, "dur": 5.785,
    "args": {
      "External id": 1180,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904192.792, "dur": 5.834,
    "args": {
      "External id": 1181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904225.208, "dur": 6.757,
    "args": {
      "External id": 1182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904226.190, "dur": 5.461,
    "args": {
      "External id": 1183,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904235.478, "dur": 8.879,
    "args": {
      "External id": 1184,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904236.590, "dur": 1.642,
    "args": {
      "External id": 1185,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904238.755, "dur": 5.450,
    "args": {
      "External id": 1186,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904265.918, "dur": 0.155,
    "args": {
      "External id": 1187,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904271.529, "dur": 78.601,
    "args": {
      "External id": 1188,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904279.028, "dur": 2.175,
    "args": {
      "External id": 1189,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904280.323, "dur": 0.417,
    "args": {
      "External id": 1190,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904282.465, "dur": 1.432,
    "args": {
      "External id": 1191,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904283.291, "dur": 0.278,
    "args": {
      "External id": 1192,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904286.087, "dur": 5.973,
    "args": {
      "External id": 1193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904294.263, "dur": 6.602,
    "args": {
      "External id": 1194,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904303.065, "dur": 6.108,
    "args": {
      "External id": 1195,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904310.526, "dur": 5.792,
    "args": {
      "External id": 1196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904317.928, "dur": 6.258,
    "args": {
      "External id": 1197,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904326.458, "dur": 6.111,
    "args": {
      "External id": 1198,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904327.160, "dur": 0.087,
    "args": {
      "External id": 1199,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904334.142, "dur": 6.386,
    "args": {
      "External id": 1200,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904341.512, "dur": 5.408,
    "args": {
      "External id": 1201,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904372.534, "dur": 5.729,
    "args": {
      "External id": 1202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904373.511, "dur": 4.518,
    "args": {
      "External id": 1203,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904381.757, "dur": 8.290,
    "args": {
      "External id": 1204,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904382.772, "dur": 1.493,
    "args": {
      "External id": 1205,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904384.832, "dur": 5.063,
    "args": {
      "External id": 1206,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904412.949, "dur": 0.176,
    "args": {
      "External id": 1207,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904418.762, "dur": 75.973,
    "args": {
      "External id": 1208,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904425.601, "dur": 2.581,
    "args": {
      "External id": 1209,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904427.190, "dur": 0.423,
    "args": {
      "External id": 1210,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904429.495, "dur": 1.462,
    "args": {
      "External id": 1211,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904430.342, "dur": 0.291,
    "args": {
      "External id": 1212,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904433.365, "dur": 6.102,
    "args": {
      "External id": 1213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904441.583, "dur": 6.154,
    "args": {
      "External id": 1214,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904449.879, "dur": 5.843,
    "args": {
      "External id": 1215,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904457.115, "dur": 5.375,
    "args": {
      "External id": 1216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904464.081, "dur": 5.701,
    "args": {
      "External id": 1217,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904471.974, "dur": 6.229,
    "args": {
      "External id": 1218,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904472.652, "dur": 0.101,
    "args": {
      "External id": 1219,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904479.708, "dur": 5.556,
    "args": {
      "External id": 1220,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904486.291, "dur": 5.293,
    "args": {
      "External id": 1221,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904514.979, "dur": 5.773,
    "args": {
      "External id": 1222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904515.849, "dur": 4.646,
    "args": {
      "External id": 1223,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904524.212, "dur": 8.375,
    "args": {
      "External id": 1224,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904525.248, "dur": 1.482,
    "args": {
      "External id": 1225,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904527.304, "dur": 5.136,
    "args": {
      "External id": 1226,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904553.339, "dur": 0.153,
    "args": {
      "External id": 1227,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904558.997, "dur": 79.629,
    "args": {
      "External id": 1228,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904565.944, "dur": 2.193,
    "args": {
      "External id": 1229,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904567.086, "dur": 0.554,
    "args": {
      "External id": 1230,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904569.371, "dur": 1.573,
    "args": {
      "External id": 1231,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904570.195, "dur": 0.376,
    "args": {
      "External id": 1232,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904573.100, "dur": 6.916,
    "args": {
      "External id": 1233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904582.324, "dur": 6.583,
    "args": {
      "External id": 1234,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904591.058, "dur": 5.821,
    "args": {
      "External id": 1235,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904598.195, "dur": 5.477,
    "args": {
      "External id": 1236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904605.371, "dur": 6.776,
    "args": {
      "External id": 1237,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904614.368, "dur": 6.505,
    "args": {
      "External id": 1238,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904614.990, "dur": 0.097,
    "args": {
      "External id": 1239,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904622.438, "dur": 5.858,
    "args": {
      "External id": 1240,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904629.256, "dur": 6.226,
    "args": {
      "External id": 1241,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904660.030, "dur": 6.052,
    "args": {
      "External id": 1242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904660.857, "dur": 5.016,
    "args": {
      "External id": 1243,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904669.648, "dur": 8.955,
    "args": {
      "External id": 1244,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904670.591, "dur": 2.158,
    "args": {
      "External id": 1245,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904673.255, "dur": 5.196,
    "args": {
      "External id": 1246,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904699.847, "dur": 0.178,
    "args": {
      "External id": 1247,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904705.944, "dur": 80.952,
    "args": {
      "External id": 1248,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904713.122, "dur": 2.795,
    "args": {
      "External id": 1249,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904714.870, "dur": 0.517,
    "args": {
      "External id": 1250,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904717.226, "dur": 1.400,
    "args": {
      "External id": 1251,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904718.011, "dur": 0.288,
    "args": {
      "External id": 1252,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904720.782, "dur": 6.481,
    "args": {
      "External id": 1253,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904729.499, "dur": 6.430,
    "args": {
      "External id": 1254,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904738.226, "dur": 6.660,
    "args": {
      "External id": 1255,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904746.228, "dur": 5.935,
    "args": {
      "External id": 1256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904753.749, "dur": 6.522,
    "args": {
      "External id": 1257,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904762.673, "dur": 6.453,
    "args": {
      "External id": 1258,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904763.247, "dur": 0.099,
    "args": {
      "External id": 1259,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904770.698, "dur": 6.391,
    "args": {
      "External id": 1260,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904778.172, "dur": 5.467,
    "args": {
      "External id": 1261,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904806.990, "dur": 6.517,
    "args": {
      "External id": 1262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904807.973, "dur": 5.323,
    "args": {
      "External id": 1263,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904817.010, "dur": 8.074,
    "args": {
      "External id": 1264,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904817.932, "dur": 1.507,
    "args": {
      "External id": 1265,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904819.921, "dur": 5.003,
    "args": {
      "External id": 1266,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904855.601, "dur": 0.410,
    "args": {
      "External id": 1267,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904863.315, "dur": 89.187,
    "args": {
      "External id": 1268,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904872.040, "dur": 3.094,
    "args": {
      "External id": 1269,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904873.777, "dur": 0.666,
    "args": {
      "External id": 1270,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740904876.546, "dur": 1.648,
    "args": {
      "External id": 1271,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904877.464, "dur": 0.389,
    "args": {
      "External id": 1272,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740904880.490, "dur": 8.081,
    "args": {
      "External id": 1273,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904891.806, "dur": 7.444,
    "args": {
      "External id": 1274,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904901.981, "dur": 6.865,
    "args": {
      "External id": 1275,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740904910.329, "dur": 6.332,
    "args": {
      "External id": 1276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740904918.667, "dur": 6.296,
    "args": {
      "External id": 1277,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740904927.455, "dur": 6.736,
    "args": {
      "External id": 1278,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740904928.235, "dur": 0.093,
    "args": {
      "External id": 1279,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740904936.013, "dur": 6.260,
    "args": {
      "External id": 1280,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740904943.399, "dur": 5.336,
    "args": {
      "External id": 1281,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740904974.935, "dur": 7.332,
    "args": {
      "External id": 1282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740904975.960, "dur": 6.032,
    "args": {
      "External id": 1283,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740904986.477, "dur": 9.310,
    "args": {
      "External id": 1284,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740904987.491, "dur": 1.630,
    "args": {
      "External id": 1285,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740904989.697, "dur": 5.943,
    "args": {
      "External id": 1286,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905017.630, "dur": 0.152,
    "args": {
      "External id": 1287,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905024.191, "dur": 81.879,
    "args": {
      "External id": 1288,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905031.874, "dur": 2.465,
    "args": {
      "External id": 1289,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905033.300, "dur": 0.563,
    "args": {
      "External id": 1290,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905035.707, "dur": 1.488,
    "args": {
      "External id": 1291,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905036.581, "dur": 0.285,
    "args": {
      "External id": 1292,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905039.399, "dur": 6.298,
    "args": {
      "External id": 1293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905048.340, "dur": 6.474,
    "args": {
      "External id": 1294,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905057.109, "dur": 6.607,
    "args": {
      "External id": 1295,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905065.050, "dur": 7.012,
    "args": {
      "External id": 1296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905073.815, "dur": 5.988,
    "args": {
      "External id": 1297,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905082.069, "dur": 6.520,
    "args": {
      "External id": 1298,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905082.757, "dur": 0.094,
    "args": {
      "External id": 1299,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905090.265, "dur": 5.746,
    "args": {
      "External id": 1300,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905097.026, "dur": 5.616,
    "args": {
      "External id": 1301,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905129.074, "dur": 6.188,
    "args": {
      "External id": 1302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905129.994, "dur": 5.016,
    "args": {
      "External id": 1303,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905138.951, "dur": 8.329,
    "args": {
      "External id": 1304,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905139.999, "dur": 1.399,
    "args": {
      "External id": 1305,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905141.910, "dur": 5.213,
    "args": {
      "External id": 1306,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905168.479, "dur": 0.156,
    "args": {
      "External id": 1307,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905174.429, "dur": 86.170,
    "args": {
      "External id": 1308,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905182.033, "dur": 2.316,
    "args": {
      "External id": 1309,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905183.452, "dur": 0.420,
    "args": {
      "External id": 1310,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905185.629, "dur": 1.426,
    "args": {
      "External id": 1311,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905186.425, "dur": 0.300,
    "args": {
      "External id": 1312,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905189.199, "dur": 6.764,
    "args": {
      "External id": 1313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905198.316, "dur": 6.345,
    "args": {
      "External id": 1314,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905206.839, "dur": 5.821,
    "args": {
      "External id": 1315,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905214.029, "dur": 5.331,
    "args": {
      "External id": 1316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905221.149, "dur": 11.330,
    "args": {
      "External id": 1317,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905234.809, "dur": 7.840,
    "args": {
      "External id": 1318,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905235.488, "dur": 0.092,
    "args": {
      "External id": 1319,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905244.283, "dur": 6.515,
    "args": {
      "External id": 1320,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905251.860, "dur": 5.615,
    "args": {
      "External id": 1321,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905281.643, "dur": 7.353,
    "args": {
      "External id": 1322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905282.565, "dur": 6.172,
    "args": {
      "External id": 1323,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905292.329, "dur": 7.871,
    "args": {
      "External id": 1324,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905293.231, "dur": 1.387,
    "args": {
      "External id": 1325,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905295.119, "dur": 4.950,
    "args": {
      "External id": 1326,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905320.097, "dur": 0.153,
    "args": {
      "External id": 1327,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905325.235, "dur": 75.870,
    "args": {
      "External id": 1328,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905331.968, "dur": 1.961,
    "args": {
      "External id": 1329,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905333.064, "dur": 0.382,
    "args": {
      "External id": 1330,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905335.155, "dur": 1.258,
    "args": {
      "External id": 1331,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905335.880, "dur": 0.260,
    "args": {
      "External id": 1332,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905338.314, "dur": 5.947,
    "args": {
      "External id": 1333,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905346.284, "dur": 6.487,
    "args": {
      "External id": 1334,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905354.699, "dur": 6.021,
    "args": {
      "External id": 1335,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905362.032, "dur": 5.235,
    "args": {
      "External id": 1336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905368.825, "dur": 7.044,
    "args": {
      "External id": 1337,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905377.981, "dur": 6.594,
    "args": {
      "External id": 1338,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905378.749, "dur": 0.086,
    "args": {
      "External id": 1339,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905386.041, "dur": 5.640,
    "args": {
      "External id": 1340,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905392.832, "dur": 5.391,
    "args": {
      "External id": 1341,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905420.490, "dur": 7.441,
    "args": {
      "External id": 1342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905421.347, "dur": 6.373,
    "args": {
      "External id": 1343,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905431.272, "dur": 8.866,
    "args": {
      "External id": 1344,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905432.364, "dur": 1.321,
    "args": {
      "External id": 1345,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905434.135, "dur": 5.864,
    "args": {
      "External id": 1346,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905459.807, "dur": 0.138,
    "args": {
      "External id": 1347,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905465.254, "dur": 75.694,
    "args": {
      "External id": 1348,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905472.031, "dur": 2.190,
    "args": {
      "External id": 1349,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905473.273, "dur": 0.510,
    "args": {
      "External id": 1350,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905475.427, "dur": 2.188,
    "args": {
      "External id": 1351,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905476.782, "dur": 0.503,
    "args": {
      "External id": 1352,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905479.496, "dur": 6.128,
    "args": {
      "External id": 1353,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905487.674, "dur": 6.341,
    "args": {
      "External id": 1354,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905496.019, "dur": 5.758,
    "args": {
      "External id": 1355,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905503.041, "dur": 6.387,
    "args": {
      "External id": 1356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905511.055, "dur": 5.722,
    "args": {
      "External id": 1357,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905518.751, "dur": 6.029,
    "args": {
      "External id": 1358,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905519.413, "dur": 0.090,
    "args": {
      "External id": 1359,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905526.230, "dur": 5.965,
    "args": {
      "External id": 1360,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905533.128, "dur": 4.852,
    "args": {
      "External id": 1361,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905559.710, "dur": 6.484,
    "args": {
      "External id": 1362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905560.523, "dur": 5.412,
    "args": {
      "External id": 1363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905569.506, "dur": 7.589,
    "args": {
      "External id": 1364,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905570.422, "dur": 1.336,
    "args": {
      "External id": 1365,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905572.253, "dur": 4.713,
    "args": {
      "External id": 1366,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905596.517, "dur": 0.160,
    "args": {
      "External id": 1367,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905601.746, "dur": 70.825,
    "args": {
      "External id": 1368,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905608.063, "dur": 1.912,
    "args": {
      "External id": 1369,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905609.197, "dur": 0.359,
    "args": {
      "External id": 1370,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905611.105, "dur": 1.258,
    "args": {
      "External id": 1371,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905611.832, "dur": 0.244,
    "args": {
      "External id": 1372,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905614.205, "dur": 5.605,
    "args": {
      "External id": 1373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905621.914, "dur": 5.834,
    "args": {
      "External id": 1374,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905629.670, "dur": 5.343,
    "args": {
      "External id": 1375,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905636.212, "dur": 4.977,
    "args": {
      "External id": 1376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905642.659, "dur": 5.599,
    "args": {
      "External id": 1377,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905650.175, "dur": 6.933,
    "args": {
      "External id": 1378,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905651.816, "dur": 0.091,
    "args": {
      "External id": 1379,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905658.532, "dur": 5.206,
    "args": {
      "External id": 1380,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905664.694, "dur": 5.056,
    "args": {
      "External id": 1381,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905691.883, "dur": 5.925,
    "args": {
      "External id": 1382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905692.704, "dur": 4.868,
    "args": {
      "External id": 1383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905701.065, "dur": 7.610,
    "args": {
      "External id": 1384,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905701.936, "dur": 1.337,
    "args": {
      "External id": 1385,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905703.730, "dur": 4.800,
    "args": {
      "External id": 1386,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905729.873, "dur": 0.167,
    "args": {
      "External id": 1387,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905735.351, "dur": 70.307,
    "args": {
      "External id": 1388,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905741.707, "dur": 1.913,
    "args": {
      "External id": 1389,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905742.811, "dur": 0.369,
    "args": {
      "External id": 1390,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905744.877, "dur": 1.322,
    "args": {
      "External id": 1391,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905745.625, "dur": 0.284,
    "args": {
      "External id": 1392,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905748.091, "dur": 5.732,
    "args": {
      "External id": 1393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905755.808, "dur": 5.845,
    "args": {
      "External id": 1394,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905763.713, "dur": 5.640,
    "args": {
      "External id": 1395,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905770.665, "dur": 5.168,
    "args": {
      "External id": 1396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905777.367, "dur": 5.355,
    "args": {
      "External id": 1397,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905784.666, "dur": 5.833,
    "args": {
      "External id": 1398,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905785.314, "dur": 0.114,
    "args": {
      "External id": 1399,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905791.880, "dur": 5.122,
    "args": {
      "External id": 1400,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905797.973, "dur": 4.724,
    "args": {
      "External id": 1401,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905824.359, "dur": 6.380,
    "args": {
      "External id": 1402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905825.200, "dur": 5.303,
    "args": {
      "External id": 1403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905834.077, "dur": 9.011,
    "args": {
      "External id": 1404,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905835.728, "dur": 1.346,
    "args": {
      "External id": 1405,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905837.651, "dur": 5.281,
    "args": {
      "External id": 1406,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905862.251, "dur": 0.238,
    "args": {
      "External id": 1407,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905867.611, "dur": 72.710,
    "args": {
      "External id": 1408,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905874.109, "dur": 2.177,
    "args": {
      "External id": 1409,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905875.385, "dur": 0.444,
    "args": {
      "External id": 1410,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740905877.429, "dur": 1.449,
    "args": {
      "External id": 1411,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905878.289, "dur": 0.295,
    "args": {
      "External id": 1412,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740905880.923, "dur": 6.089,
    "args": {
      "External id": 1413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905889.018, "dur": 6.214,
    "args": {
      "External id": 1414,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905897.159, "dur": 5.406,
    "args": {
      "External id": 1415,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740905903.768, "dur": 5.133,
    "args": {
      "External id": 1416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740905910.379, "dur": 6.317,
    "args": {
      "External id": 1417,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740905918.703, "dur": 6.012,
    "args": {
      "External id": 1418,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905919.300, "dur": 0.076,
    "args": {
      "External id": 1419,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740905926.071, "dur": 5.426,
    "args": {
      "External id": 1420,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740905932.437, "dur": 4.897,
    "args": {
      "External id": 1421,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740905959.397, "dur": 7.095,
    "args": {
      "External id": 1422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740905961.347, "dur": 4.946,
    "args": {
      "External id": 1423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740905969.994, "dur": 7.982,
    "args": {
      "External id": 1424,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740905971.006, "dur": 1.314,
    "args": {
      "External id": 1425,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740905972.756, "dur": 5.077,
    "args": {
      "External id": 1426,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740905997.461, "dur": 0.179,
    "args": {
      "External id": 1427,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906002.532, "dur": 85.784,
    "args": {
      "External id": 1428,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906008.902, "dur": 1.960,
    "args": {
      "External id": 1429,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906010.028, "dur": 0.403,
    "args": {
      "External id": 1430,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906012.038, "dur": 1.311,
    "args": {
      "External id": 1431,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906012.792, "dur": 0.255,
    "args": {
      "External id": 1432,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906015.310, "dur": 5.674,
    "args": {
      "External id": 1433,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906022.867, "dur": 5.711,
    "args": {
      "External id": 1434,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906030.477, "dur": 5.140,
    "args": {
      "External id": 1435,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906036.934, "dur": 5.239,
    "args": {
      "External id": 1436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906043.597, "dur": 5.651,
    "args": {
      "External id": 1437,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906051.300, "dur": 5.912,
    "args": {
      "External id": 1438,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906051.870, "dur": 0.084,
    "args": {
      "External id": 1439,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906058.652, "dur": 5.769,
    "args": {
      "External id": 1440,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906065.362, "dur": 18.128,
    "args": {
      "External id": 1441,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906111.401, "dur": 8.571,
    "args": {
      "External id": 1442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906112.685, "dur": 7.009,
    "args": {
      "External id": 1443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906123.655, "dur": 9.968,
    "args": {
      "External id": 1444,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906124.825, "dur": 1.837,
    "args": {
      "External id": 1445,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906127.203, "dur": 6.291,
    "args": {
      "External id": 1446,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906155.026, "dur": 0.188,
    "args": {
      "External id": 1447,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906161.011, "dur": 85.062,
    "args": {
      "External id": 1448,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906168.768, "dur": 5.399,
    "args": {
      "External id": 1449,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906172.853, "dur": 0.756,
    "args": {
      "External id": 1450,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906175.401, "dur": 1.975,
    "args": {
      "External id": 1451,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906176.468, "dur": 0.614,
    "args": {
      "External id": 1452,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906179.706, "dur": 6.541,
    "args": {
      "External id": 1453,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906188.804, "dur": 6.740,
    "args": {
      "External id": 1454,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906197.606, "dur": 5.598,
    "args": {
      "External id": 1455,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906204.584, "dur": 6.468,
    "args": {
      "External id": 1456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906212.601, "dur": 6.319,
    "args": {
      "External id": 1457,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906221.054, "dur": 6.889,
    "args": {
      "External id": 1458,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906221.759, "dur": 0.104,
    "args": {
      "External id": 1459,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906229.563, "dur": 7.112,
    "args": {
      "External id": 1460,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906237.649, "dur": 5.305,
    "args": {
      "External id": 1461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906267.301, "dur": 6.653,
    "args": {
      "External id": 1462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906268.120, "dur": 5.618,
    "args": {
      "External id": 1463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906277.381, "dur": 8.618,
    "args": {
      "External id": 1464,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906278.270, "dur": 1.340,
    "args": {
      "External id": 1465,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906280.088, "dur": 5.780,
    "args": {
      "External id": 1466,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906306.118, "dur": 0.133,
    "args": {
      "External id": 1467,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906311.674, "dur": 70.908,
    "args": {
      "External id": 1468,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906318.702, "dur": 2.016,
    "args": {
      "External id": 1469,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906319.753, "dur": 0.431,
    "args": {
      "External id": 1470,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906321.963, "dur": 1.283,
    "args": {
      "External id": 1471,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906322.716, "dur": 0.248,
    "args": {
      "External id": 1472,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906325.348, "dur": 5.948,
    "args": {
      "External id": 1473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906333.448, "dur": 5.795,
    "args": {
      "External id": 1474,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906341.125, "dur": 4.967,
    "args": {
      "External id": 1475,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906347.309, "dur": 5.053,
    "args": {
      "External id": 1476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906353.816, "dur": 5.459,
    "args": {
      "External id": 1477,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906361.217, "dur": 5.597,
    "args": {
      "External id": 1478,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906361.803, "dur": 0.085,
    "args": {
      "External id": 1479,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906368.294, "dur": 5.332,
    "args": {
      "External id": 1480,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906374.594, "dur": 4.941,
    "args": {
      "External id": 1481,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906403.267, "dur": 5.570,
    "args": {
      "External id": 1482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906404.036, "dur": 4.541,
    "args": {
      "External id": 1483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906412.206, "dur": 10.160,
    "args": {
      "External id": 1484,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906413.186, "dur": 1.335,
    "args": {
      "External id": 1485,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906414.999, "dur": 7.236,
    "args": {
      "External id": 1486,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906442.232, "dur": 0.204,
    "args": {
      "External id": 1487,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906448.336, "dur": 82.660,
    "args": {
      "External id": 1488,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906455.294, "dur": 2.076,
    "args": {
      "External id": 1489,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906456.468, "dur": 0.415,
    "args": {
      "External id": 1490,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906458.523, "dur": 1.440,
    "args": {
      "External id": 1491,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906459.273, "dur": 0.396,
    "args": {
      "External id": 1492,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906461.958, "dur": 6.120,
    "args": {
      "External id": 1493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906474.498, "dur": 6.321,
    "args": {
      "External id": 1494,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906482.828, "dur": 5.352,
    "args": {
      "External id": 1495,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906489.465, "dur": 5.500,
    "args": {
      "External id": 1496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906496.435, "dur": 5.763,
    "args": {
      "External id": 1497,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906504.219, "dur": 10.273,
    "args": {
      "External id": 1498,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906504.814, "dur": 0.085,
    "args": {
      "External id": 1499,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906516.063, "dur": 5.424,
    "args": {
      "External id": 1500,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906522.481, "dur": 5.278,
    "args": {
      "External id": 1501,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906552.405, "dur": 6.794,
    "args": {
      "External id": 1502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906553.275, "dur": 5.710,
    "args": {
      "External id": 1503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906562.540, "dur": 8.757,
    "args": {
      "External id": 1504,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906563.539, "dur": 1.626,
    "args": {
      "External id": 1505,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906565.618, "dur": 5.556,
    "args": {
      "External id": 1506,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906591.218, "dur": 0.178,
    "args": {
      "External id": 1507,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906596.686, "dur": 76.414,
    "args": {
      "External id": 1508,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906603.549, "dur": 2.133,
    "args": {
      "External id": 1509,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906604.749, "dur": 0.357,
    "args": {
      "External id": 1510,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906606.865, "dur": 1.350,
    "args": {
      "External id": 1511,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906607.618, "dur": 0.270,
    "args": {
      "External id": 1512,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906610.202, "dur": 5.956,
    "args": {
      "External id": 1513,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906618.204, "dur": 5.955,
    "args": {
      "External id": 1514,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906626.140, "dur": 7.062,
    "args": {
      "External id": 1515,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906634.608, "dur": 5.753,
    "args": {
      "External id": 1516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906641.890, "dur": 5.519,
    "args": {
      "External id": 1517,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906649.367, "dur": 7.094,
    "args": {
      "External id": 1518,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906649.935, "dur": 0.083,
    "args": {
      "External id": 1519,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906657.877, "dur": 5.963,
    "args": {
      "External id": 1520,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906664.816, "dur": 5.214,
    "args": {
      "External id": 1521,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906691.216, "dur": 5.897,
    "args": {
      "External id": 1522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906692.090, "dur": 4.762,
    "args": {
      "External id": 1523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906700.527, "dur": 7.940,
    "args": {
      "External id": 1524,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906701.491, "dur": 1.358,
    "args": {
      "External id": 1525,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906703.342, "dur": 4.988,
    "args": {
      "External id": 1526,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906728.078, "dur": 0.173,
    "args": {
      "External id": 1527,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906733.592, "dur": 97.180,
    "args": {
      "External id": 1528,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906740.424, "dur": 1.907,
    "args": {
      "External id": 1529,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906741.476, "dur": 0.405,
    "args": {
      "External id": 1530,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906743.569, "dur": 1.295,
    "args": {
      "External id": 1531,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906744.308, "dur": 0.250,
    "args": {
      "External id": 1532,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906746.854, "dur": 6.004,
    "args": {
      "External id": 1533,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906755.009, "dur": 5.927,
    "args": {
      "External id": 1534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906762.944, "dur": 6.451,
    "args": {
      "External id": 1535,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906770.832, "dur": 5.352,
    "args": {
      "External id": 1536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906798.617, "dur": 5.799,
    "args": {
      "External id": 1537,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906806.412, "dur": 6.887,
    "args": {
      "External id": 1538,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906807.006, "dur": 0.087,
    "args": {
      "External id": 1539,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906814.708, "dur": 5.830,
    "args": {
      "External id": 1540,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906821.445, "dur": 6.397,
    "args": {
      "External id": 1541,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906849.409, "dur": 6.035,
    "args": {
      "External id": 1542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906850.378, "dur": 4.797,
    "args": {
      "External id": 1543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906858.828, "dur": 7.792,
    "args": {
      "External id": 1544,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906859.680, "dur": 1.320,
    "args": {
      "External id": 1545,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906861.595, "dur": 4.890,
    "args": {
      "External id": 1546,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906886.124, "dur": 0.158,
    "args": {
      "External id": 1547,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906891.408, "dur": 76.096,
    "args": {
      "External id": 1548,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906897.853, "dur": 1.992,
    "args": {
      "External id": 1549,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906898.930, "dur": 0.395,
    "args": {
      "External id": 1550,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740906901.089, "dur": 1.316,
    "args": {
      "External id": 1551,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906901.848, "dur": 0.251,
    "args": {
      "External id": 1552,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740906904.429, "dur": 5.984,
    "args": {
      "External id": 1553,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906912.381, "dur": 6.152,
    "args": {
      "External id": 1554,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906920.513, "dur": 6.759,
    "args": {
      "External id": 1555,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740906928.599, "dur": 5.570,
    "args": {
      "External id": 1556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740906935.686, "dur": 5.788,
    "args": {
      "External id": 1557,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740906943.358, "dur": 7.334,
    "args": {
      "External id": 1558,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740906943.948, "dur": 0.081,
    "args": {
      "External id": 1559,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740906952.097, "dur": 5.662,
    "args": {
      "External id": 1560,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740906958.706, "dur": 5.803,
    "args": {
      "External id": 1561,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740906986.123, "dur": 6.086,
    "args": {
      "External id": 1562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740906987.047, "dur": 4.906,
    "args": {
      "External id": 1563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740906995.530, "dur": 8.055,
    "args": {
      "External id": 1564,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740906996.336, "dur": 1.337,
    "args": {
      "External id": 1565,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740906998.105, "dur": 5.343,
    "args": {
      "External id": 1566,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907024.373, "dur": 0.193,
    "args": {
      "External id": 1567,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907030.329, "dur": 95.398,
    "args": {
      "External id": 1568,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907036.986, "dur": 2.105,
    "args": {
      "External id": 1569,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907038.179, "dur": 0.414,
    "args": {
      "External id": 1570,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907040.269, "dur": 1.375,
    "args": {
      "External id": 1571,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907040.995, "dur": 0.346,
    "args": {
      "External id": 1572,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907044.776, "dur": 6.366,
    "args": {
      "External id": 1573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907053.273, "dur": 6.482,
    "args": {
      "External id": 1574,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907061.681, "dur": 7.186,
    "args": {
      "External id": 1575,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907070.183, "dur": 6.359,
    "args": {
      "External id": 1576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907077.928, "dur": 19.816,
    "args": {
      "External id": 1577,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907099.847, "dur": 8.205,
    "args": {
      "External id": 1578,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907100.512, "dur": 0.094,
    "args": {
      "External id": 1579,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907109.575, "dur": 5.855,
    "args": {
      "External id": 1580,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907116.407, "dur": 6.188,
    "args": {
      "External id": 1581,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907144.935, "dur": 7.001,
    "args": {
      "External id": 1582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907145.783, "dur": 5.898,
    "args": {
      "External id": 1583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907155.132, "dur": 8.769,
    "args": {
      "External id": 1584,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907156.051, "dur": 1.582,
    "args": {
      "External id": 1585,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907158.095, "dur": 5.671,
    "args": {
      "External id": 1586,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907184.122, "dur": 0.141,
    "args": {
      "External id": 1587,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907189.315, "dur": 76.739,
    "args": {
      "External id": 1588,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907195.963, "dur": 2.028,
    "args": {
      "External id": 1589,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907197.070, "dur": 0.384,
    "args": {
      "External id": 1590,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907199.193, "dur": 1.637,
    "args": {
      "External id": 1591,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907200.010, "dur": 0.497,
    "args": {
      "External id": 1592,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907202.837, "dur": 6.368,
    "args": {
      "External id": 1593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907211.472, "dur": 6.035,
    "args": {
      "External id": 1594,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907219.550, "dur": 6.377,
    "args": {
      "External id": 1595,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907227.292, "dur": 5.596,
    "args": {
      "External id": 1596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907234.404, "dur": 6.176,
    "args": {
      "External id": 1597,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907242.626, "dur": 6.937,
    "args": {
      "External id": 1598,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907243.205, "dur": 0.087,
    "args": {
      "External id": 1599,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907250.974, "dur": 6.218,
    "args": {
      "External id": 1600,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907258.165, "dur": 4.928,
    "args": {
      "External id": 1601,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907285.607, "dur": 6.008,
    "args": {
      "External id": 1602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907286.468, "dur": 4.927,
    "args": {
      "External id": 1603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907295.094, "dur": 8.000,
    "args": {
      "External id": 1604,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907295.934, "dur": 1.267,
    "args": {
      "External id": 1605,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907297.637, "dur": 5.317,
    "args": {
      "External id": 1606,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907322.527, "dur": 0.112,
    "args": {
      "External id": 1607,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907327.425, "dur": 75.267,
    "args": {
      "External id": 1608,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907333.815, "dur": 2.031,
    "args": {
      "External id": 1609,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907334.938, "dur": 0.400,
    "args": {
      "External id": 1610,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907337.037, "dur": 1.299,
    "args": {
      "External id": 1611,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907337.764, "dur": 0.250,
    "args": {
      "External id": 1612,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907340.216, "dur": 5.626,
    "args": {
      "External id": 1613,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907347.803, "dur": 5.991,
    "args": {
      "External id": 1614,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907355.757, "dur": 6.949,
    "args": {
      "External id": 1615,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907364.005, "dur": 5.364,
    "args": {
      "External id": 1616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907371.709, "dur": 5.751,
    "args": {
      "External id": 1617,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907379.465, "dur": 6.900,
    "args": {
      "External id": 1618,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907380.004, "dur": 0.087,
    "args": {
      "External id": 1619,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907387.811, "dur": 5.897,
    "args": {
      "External id": 1620,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907394.659, "dur": 5.035,
    "args": {
      "External id": 1621,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907421.045, "dur": 5.922,
    "args": {
      "External id": 1622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907421.941, "dur": 4.775,
    "args": {
      "External id": 1623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907430.235, "dur": 7.536,
    "args": {
      "External id": 1624,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907431.037, "dur": 1.330,
    "args": {
      "External id": 1625,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907432.822, "dur": 4.826,
    "args": {
      "External id": 1626,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907457.091, "dur": 0.141,
    "args": {
      "External id": 1627,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907462.210, "dur": 73.438,
    "args": {
      "External id": 1628,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907468.644, "dur": 1.944,
    "args": {
      "External id": 1629,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907469.699, "dur": 0.387,
    "args": {
      "External id": 1630,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907471.795, "dur": 1.331,
    "args": {
      "External id": 1631,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907472.563, "dur": 0.246,
    "args": {
      "External id": 1632,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907475.083, "dur": 5.755,
    "args": {
      "External id": 1633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907482.765, "dur": 5.854,
    "args": {
      "External id": 1634,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907490.452, "dur": 6.511,
    "args": {
      "External id": 1635,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907498.329, "dur": 5.462,
    "args": {
      "External id": 1636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907505.267, "dur": 5.789,
    "args": {
      "External id": 1637,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907513.060, "dur": 6.618,
    "args": {
      "External id": 1638,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907513.640, "dur": 0.081,
    "args": {
      "External id": 1639,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907521.105, "dur": 5.599,
    "args": {
      "External id": 1640,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907527.654, "dur": 5.008,
    "args": {
      "External id": 1641,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907554.192, "dur": 5.925,
    "args": {
      "External id": 1642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907555.029, "dur": 4.818,
    "args": {
      "External id": 1643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907563.446, "dur": 7.721,
    "args": {
      "External id": 1644,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907564.192, "dur": 1.411,
    "args": {
      "External id": 1645,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907566.051, "dur": 4.991,
    "args": {
      "External id": 1646,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907590.591, "dur": 0.128,
    "args": {
      "External id": 1647,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907595.416, "dur": 80.514,
    "args": {
      "External id": 1648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907601.665, "dur": 1.952,
    "args": {
      "External id": 1649,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907602.744, "dur": 0.383,
    "args": {
      "External id": 1650,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907604.775, "dur": 1.313,
    "args": {
      "External id": 1651,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907605.505, "dur": 0.260,
    "args": {
      "External id": 1652,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907607.901, "dur": 5.553,
    "args": {
      "External id": 1653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907615.239, "dur": 5.860,
    "args": {
      "External id": 1654,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907622.934, "dur": 9.432,
    "args": {
      "External id": 1655,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907633.898, "dur": 6.272,
    "args": {
      "External id": 1656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907641.717, "dur": 7.213,
    "args": {
      "External id": 1657,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907650.995, "dur": 7.715,
    "args": {
      "External id": 1658,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907651.593, "dur": 0.077,
    "args": {
      "External id": 1659,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907660.090, "dur": 6.444,
    "args": {
      "External id": 1660,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907667.453, "dur": 5.562,
    "args": {
      "External id": 1661,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907694.888, "dur": 6.368,
    "args": {
      "External id": 1662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907695.760, "dur": 5.292,
    "args": {
      "External id": 1663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907704.453, "dur": 8.578,
    "args": {
      "External id": 1664,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907705.218, "dur": 1.304,
    "args": {
      "External id": 1665,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907706.996, "dur": 5.906,
    "args": {
      "External id": 1666,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907734.272, "dur": 0.151,
    "args": {
      "External id": 1667,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907739.532, "dur": 77.697,
    "args": {
      "External id": 1668,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907746.041, "dur": 2.049,
    "args": {
      "External id": 1669,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907747.170, "dur": 0.409,
    "args": {
      "External id": 1670,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907749.263, "dur": 1.459,
    "args": {
      "External id": 1671,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907750.019, "dur": 0.406,
    "args": {
      "External id": 1672,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907752.678, "dur": 6.938,
    "args": {
      "External id": 1673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907761.504, "dur": 6.399,
    "args": {
      "External id": 1674,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907769.887, "dur": 7.153,
    "args": {
      "External id": 1675,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907778.330, "dur": 5.611,
    "args": {
      "External id": 1676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907785.372, "dur": 6.042,
    "args": {
      "External id": 1677,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907793.394, "dur": 7.080,
    "args": {
      "External id": 1678,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907793.989, "dur": 0.077,
    "args": {
      "External id": 1679,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907801.884, "dur": 6.184,
    "args": {
      "External id": 1680,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907808.971, "dur": 5.321,
    "args": {
      "External id": 1681,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907835.294, "dur": 6.270,
    "args": {
      "External id": 1682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907836.175, "dur": 5.181,
    "args": {
      "External id": 1683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907845.009, "dur": 7.973,
    "args": {
      "External id": 1684,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907845.917, "dur": 1.285,
    "args": {
      "External id": 1685,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907847.646, "dur": 5.209,
    "args": {
      "External id": 1686,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907872.321, "dur": 0.151,
    "args": {
      "External id": 1687,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907877.186, "dur": 77.559,
    "args": {
      "External id": 1688,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907883.698, "dur": 1.910,
    "args": {
      "External id": 1689,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907884.788, "dur": 0.344,
    "args": {
      "External id": 1690,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740907886.748, "dur": 1.263,
    "args": {
      "External id": 1691,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907887.468, "dur": 0.255,
    "args": {
      "External id": 1692,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740907889.859, "dur": 5.711,
    "args": {
      "External id": 1693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907897.679, "dur": 5.919,
    "args": {
      "External id": 1694,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907905.550, "dur": 7.177,
    "args": {
      "External id": 1695,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740907913.999, "dur": 5.739,
    "args": {
      "External id": 1696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740907921.117, "dur": 5.626,
    "args": {
      "External id": 1697,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740907928.664, "dur": 9.137,
    "args": {
      "External id": 1698,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740907929.245, "dur": 0.085,
    "args": {
      "External id": 1699,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740907939.168, "dur": 6.298,
    "args": {
      "External id": 1700,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740907946.419, "dur": 5.469,
    "args": {
      "External id": 1701,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740907973.571, "dur": 6.768,
    "args": {
      "External id": 1702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740907974.438, "dur": 5.697,
    "args": {
      "External id": 1703,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740907983.808, "dur": 8.153,
    "args": {
      "External id": 1704,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740907984.707, "dur": 1.312,
    "args": {
      "External id": 1705,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740907986.480, "dur": 5.320,
    "args": {
      "External id": 1706,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908012.687, "dur": 0.137,
    "args": {
      "External id": 1707,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908017.689, "dur": 78.995,
    "args": {
      "External id": 1708,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908024.043, "dur": 2.015,
    "args": {
      "External id": 1709,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908025.144, "dur": 0.391,
    "args": {
      "External id": 1710,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908027.170, "dur": 1.396,
    "args": {
      "External id": 1711,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908027.933, "dur": 0.343,
    "args": {
      "External id": 1712,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908030.554, "dur": 6.732,
    "args": {
      "External id": 1713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908039.416, "dur": 6.825,
    "args": {
      "External id": 1714,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908048.135, "dur": 6.098,
    "args": {
      "External id": 1715,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908055.479, "dur": 5.538,
    "args": {
      "External id": 1716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908062.556, "dur": 7.545,
    "args": {
      "External id": 1717,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908072.291, "dur": 6.163,
    "args": {
      "External id": 1718,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908072.866, "dur": 0.093,
    "args": {
      "External id": 1719,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908079.856, "dur": 6.068,
    "args": {
      "External id": 1720,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908086.916, "dur": 6.887,
    "args": {
      "External id": 1721,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908116.098, "dur": 6.527,
    "args": {
      "External id": 1722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908116.973, "dur": 5.406,
    "args": {
      "External id": 1723,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908126.112, "dur": 8.995,
    "args": {
      "External id": 1724,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908126.989, "dur": 1.339,
    "args": {
      "External id": 1725,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908128.851, "dur": 6.111,
    "args": {
      "External id": 1726,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908156.007, "dur": 0.091,
    "args": {
      "External id": 1727,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908160.838, "dur": 82.696,
    "args": {
      "External id": 1728,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908167.022, "dur": 2.163,
    "args": {
      "External id": 1729,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908168.299, "dur": 0.389,
    "args": {
      "External id": 1730,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908170.365, "dur": 1.274,
    "args": {
      "External id": 1731,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908171.090, "dur": 0.253,
    "args": {
      "External id": 1732,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908173.455, "dur": 6.231,
    "args": {
      "External id": 1733,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908181.700, "dur": 6.274,
    "args": {
      "External id": 1734,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908189.860, "dur": 5.758,
    "args": {
      "External id": 1735,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908196.840, "dur": 5.788,
    "args": {
      "External id": 1736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908204.128, "dur": 6.542,
    "args": {
      "External id": 1737,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908212.749, "dur": 5.990,
    "args": {
      "External id": 1738,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908213.309, "dur": 0.093,
    "args": {
      "External id": 1739,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908220.129, "dur": 6.616,
    "args": {
      "External id": 1740,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908227.704, "dur": 12.884,
    "args": {
      "External id": 1741,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908263.875, "dur": 7.697,
    "args": {
      "External id": 1742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908264.782, "dur": 6.485,
    "args": {
      "External id": 1743,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908274.981, "dur": 10.394,
    "args": {
      "External id": 1744,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908275.854, "dur": 1.352,
    "args": {
      "External id": 1745,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908277.691, "dur": 7.514,
    "args": {
      "External id": 1746,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908306.771, "dur": 0.115,
    "args": {
      "External id": 1747,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908312.032, "dur": 85.457,
    "args": {
      "External id": 1748,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908318.672, "dur": 2.087,
    "args": {
      "External id": 1749,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908319.882, "dur": 0.400,
    "args": {
      "External id": 1750,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908321.903, "dur": 1.447,
    "args": {
      "External id": 1751,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908322.778, "dur": 0.260,
    "args": {
      "External id": 1752,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908325.329, "dur": 6.649,
    "args": {
      "External id": 1753,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908334.420, "dur": 8.319,
    "args": {
      "External id": 1754,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908344.837, "dur": 6.502,
    "args": {
      "External id": 1755,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908352.688, "dur": 6.646,
    "args": {
      "External id": 1756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908360.938, "dur": 7.530,
    "args": {
      "External id": 1757,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908370.470, "dur": 6.319,
    "args": {
      "External id": 1758,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908371.172, "dur": 0.088,
    "args": {
      "External id": 1759,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908378.474, "dur": 7.353,
    "args": {
      "External id": 1760,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908386.791, "dur": 7.630,
    "args": {
      "External id": 1761,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908418.180, "dur": 6.915,
    "args": {
      "External id": 1762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908419.036, "dur": 5.844,
    "args": {
      "External id": 1763,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908428.590, "dur": 9.045,
    "args": {
      "External id": 1764,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908429.503, "dur": 1.358,
    "args": {
      "External id": 1765,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908431.351, "dur": 6.127,
    "args": {
      "External id": 1766,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908457.428, "dur": 0.137,
    "args": {
      "External id": 1767,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908462.430, "dur": 80.051,
    "args": {
      "External id": 1768,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908468.496, "dur": 2.938,
    "args": {
      "External id": 1769,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908470.346, "dur": 0.590,
    "args": {
      "External id": 1770,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908472.648, "dur": 1.399,
    "args": {
      "External id": 1771,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908473.505, "dur": 0.248,
    "args": {
      "External id": 1772,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908475.984, "dur": 7.444,
    "args": {
      "External id": 1773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908485.392, "dur": 7.135,
    "args": {
      "External id": 1774,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908494.577, "dur": 6.707,
    "args": {
      "External id": 1775,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908502.590, "dur": 5.843,
    "args": {
      "External id": 1776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908509.879, "dur": 5.836,
    "args": {
      "External id": 1777,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908517.706, "dur": 6.657,
    "args": {
      "External id": 1778,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908518.303, "dur": 0.086,
    "args": {
      "External id": 1779,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908525.826, "dur": 6.710,
    "args": {
      "External id": 1780,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908533.465, "dur": 6.098,
    "args": {
      "External id": 1781,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908561.259, "dur": 6.600,
    "args": {
      "External id": 1782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908562.062, "dur": 5.584,
    "args": {
      "External id": 1783,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908571.203, "dur": 8.735,
    "args": {
      "External id": 1784,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908572.026, "dur": 1.286,
    "args": {
      "External id": 1785,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908573.783, "dur": 5.980,
    "args": {
      "External id": 1786,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908599.399, "dur": 0.125,
    "args": {
      "External id": 1787,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908604.318, "dur": 77.202,
    "args": {
      "External id": 1788,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908610.744, "dur": 1.930,
    "args": {
      "External id": 1789,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908611.826, "dur": 0.349,
    "args": {
      "External id": 1790,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908613.910, "dur": 1.305,
    "args": {
      "External id": 1791,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908614.649, "dur": 0.271,
    "args": {
      "External id": 1792,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908617.176, "dur": 6.909,
    "args": {
      "External id": 1793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908626.045, "dur": 6.584,
    "args": {
      "External id": 1794,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908634.492, "dur": 7.010,
    "args": {
      "External id": 1795,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908642.910, "dur": 5.605,
    "args": {
      "External id": 1796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908650.085, "dur": 5.865,
    "args": {
      "External id": 1797,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908657.913, "dur": 6.466,
    "args": {
      "External id": 1798,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908658.486, "dur": 0.070,
    "args": {
      "External id": 1799,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908665.858, "dur": 6.310,
    "args": {
      "External id": 1800,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908673.090, "dur": 5.547,
    "args": {
      "External id": 1801,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908700.233, "dur": 6.732,
    "args": {
      "External id": 1802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908701.123, "dur": 5.626,
    "args": {
      "External id": 1803,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908710.112, "dur": 8.332,
    "args": {
      "External id": 1804,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908711.004, "dur": 1.326,
    "args": {
      "External id": 1805,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908712.747, "dur": 5.540,
    "args": {
      "External id": 1806,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908737.713, "dur": 0.129,
    "args": {
      "External id": 1807,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908742.723, "dur": 76.132,
    "args": {
      "External id": 1808,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908749.207, "dur": 1.967,
    "args": {
      "External id": 1809,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908750.305, "dur": 0.386,
    "args": {
      "External id": 1810,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908752.342, "dur": 1.317,
    "args": {
      "External id": 1811,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908753.105, "dur": 0.249,
    "args": {
      "External id": 1812,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908755.503, "dur": 6.694,
    "args": {
      "External id": 1813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908764.146, "dur": 6.617,
    "args": {
      "External id": 1814,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908772.816, "dur": 5.904,
    "args": {
      "External id": 1815,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908780.021, "dur": 5.686,
    "args": {
      "External id": 1816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908787.231, "dur": 6.677,
    "args": {
      "External id": 1817,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908795.954, "dur": 6.282,
    "args": {
      "External id": 1818,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908796.529, "dur": 0.094,
    "args": {
      "External id": 1819,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908803.570, "dur": 5.996,
    "args": {
      "External id": 1820,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908810.590, "dur": 5.454,
    "args": {
      "External id": 1821,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740908852.078, "dur": 8.476,
    "args": {
      "External id": 1822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740908853.305, "dur": 6.877,
    "args": {
      "External id": 1823,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740908864.606, "dur": 13.060,
    "args": {
      "External id": 1824,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908866.052, "dur": 2.056,
    "args": {
      "External id": 1825,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740908868.700, "dur": 8.829,
    "args": {
      "External id": 1826,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908900.755, "dur": 0.171,
    "args": {
      "External id": 1827,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908907.113, "dur": 87.385,
    "args": {
      "External id": 1828,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908914.731, "dur": 2.547,
    "args": {
      "External id": 1829,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908916.266, "dur": 0.445,
    "args": {
      "External id": 1830,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740908918.487, "dur": 1.423,
    "args": {
      "External id": 1831,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740908919.270, "dur": 0.318,
    "args": {
      "External id": 1832,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740908921.822, "dur": 8.555,
    "args": {
      "External id": 1833,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908932.884, "dur": 7.168,
    "args": {
      "External id": 1834,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908942.261, "dur": 5.560,
    "args": {
      "External id": 1835,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740908949.253, "dur": 6.908,
    "args": {
      "External id": 1836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740908957.795, "dur": 8.466,
    "args": {
      "External id": 1837,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740908968.439, "dur": 6.763,
    "args": {
      "External id": 1838,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740908969.079, "dur": 0.089,
    "args": {
      "External id": 1839,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740908976.682, "dur": 7.814,
    "args": {
      "External id": 1840,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740908985.485, "dur": 5.736,
    "args": {
      "External id": 1841,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909015.576, "dur": 7.132,
    "args": {
      "External id": 1842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909016.444, "dur": 6.049,
    "args": {
      "External id": 1843,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909026.247, "dur": 9.087,
    "args": {
      "External id": 1844,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909027.111, "dur": 1.361,
    "args": {
      "External id": 1845,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909028.935, "dur": 6.241,
    "args": {
      "External id": 1846,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909056.629, "dur": 0.153,
    "args": {
      "External id": 1847,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909062.222, "dur": 79.121,
    "args": {
      "External id": 1848,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909069.291, "dur": 1.925,
    "args": {
      "External id": 1849,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909070.323, "dur": 0.369,
    "args": {
      "External id": 1850,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909072.408, "dur": 1.323,
    "args": {
      "External id": 1851,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909073.161, "dur": 0.268,
    "args": {
      "External id": 1852,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909075.500, "dur": 7.764,
    "args": {
      "External id": 1853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909085.378, "dur": 6.615,
    "args": {
      "External id": 1854,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909094.076, "dur": 5.751,
    "args": {
      "External id": 1855,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909101.486, "dur": 5.380,
    "args": {
      "External id": 1856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909108.464, "dur": 6.315,
    "args": {
      "External id": 1857,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909116.914, "dur": 6.746,
    "args": {
      "External id": 1858,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909117.522, "dur": 0.094,
    "args": {
      "External id": 1859,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909125.232, "dur": 5.916,
    "args": {
      "External id": 1860,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909132.081, "dur": 6.097,
    "args": {
      "External id": 1861,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909160.930, "dur": 6.407,
    "args": {
      "External id": 1862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909161.724, "dur": 5.390,
    "args": {
      "External id": 1863,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909170.679, "dur": 8.641,
    "args": {
      "External id": 1864,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909171.616, "dur": 1.237,
    "args": {
      "External id": 1865,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909173.363, "dur": 5.826,
    "args": {
      "External id": 1866,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909199.262, "dur": 0.152,
    "args": {
      "External id": 1867,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909204.612, "dur": 79.728,
    "args": {
      "External id": 1868,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909211.118, "dur": 1.997,
    "args": {
      "External id": 1869,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909212.254, "dur": 0.384,
    "args": {
      "External id": 1870,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909214.365, "dur": 1.365,
    "args": {
      "External id": 1871,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909215.179, "dur": 0.244,
    "args": {
      "External id": 1872,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909217.721, "dur": 8.993,
    "args": {
      "External id": 1873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909228.681, "dur": 7.022,
    "args": {
      "External id": 1874,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909237.714, "dur": 6.114,
    "args": {
      "External id": 1875,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909245.086, "dur": 5.278,
    "args": {
      "External id": 1876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909251.935, "dur": 6.360,
    "args": {
      "External id": 1877,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909260.325, "dur": 6.704,
    "args": {
      "External id": 1878,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909260.893, "dur": 0.119,
    "args": {
      "External id": 1879,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909268.495, "dur": 6.045,
    "args": {
      "External id": 1880,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909275.445, "dur": 5.720,
    "args": {
      "External id": 1881,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909303.930, "dur": 6.334,
    "args": {
      "External id": 1882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909304.711, "dur": 5.319,
    "args": {
      "External id": 1883,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909313.645, "dur": 8.243,
    "args": {
      "External id": 1884,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909314.539, "dur": 1.305,
    "args": {
      "External id": 1885,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909316.308, "dur": 5.448,
    "args": {
      "External id": 1886,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909341.811, "dur": 0.126,
    "args": {
      "External id": 1887,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909346.889, "dur": 84.087,
    "args": {
      "External id": 1888,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909353.314, "dur": 2.063,
    "args": {
      "External id": 1889,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909354.454, "dur": 0.419,
    "args": {
      "External id": 1890,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909356.541, "dur": 1.298,
    "args": {
      "External id": 1891,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909357.285, "dur": 0.246,
    "args": {
      "External id": 1892,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909360.096, "dur": 6.769,
    "args": {
      "External id": 1893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909368.755, "dur": 6.749,
    "args": {
      "External id": 1894,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909377.571, "dur": 6.533,
    "args": {
      "External id": 1895,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909385.436, "dur": 5.874,
    "args": {
      "External id": 1896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909392.773, "dur": 6.491,
    "args": {
      "External id": 1897,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909401.343, "dur": 5.958,
    "args": {
      "External id": 1898,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909401.880, "dur": 0.095,
    "args": {
      "External id": 1899,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909408.718, "dur": 6.296,
    "args": {
      "External id": 1900,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909416.001, "dur": 11.933,
    "args": {
      "External id": 1901,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909452.063, "dur": 7.553,
    "args": {
      "External id": 1902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909453.003, "dur": 6.356,
    "args": {
      "External id": 1903,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909463.316, "dur": 8.886,
    "args": {
      "External id": 1904,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909464.184, "dur": 1.368,
    "args": {
      "External id": 1905,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909466.026, "dur": 6.048,
    "args": {
      "External id": 1906,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909493.677, "dur": 0.121,
    "args": {
      "External id": 1907,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909499.137, "dur": 78.615,
    "args": {
      "External id": 1908,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909505.536, "dur": 2.157,
    "args": {
      "External id": 1909,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909506.741, "dur": 0.461,
    "args": {
      "External id": 1910,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909508.856, "dur": 1.401,
    "args": {
      "External id": 1911,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909509.641, "dur": 0.301,
    "args": {
      "External id": 1912,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909512.078, "dur": 6.823,
    "args": {
      "External id": 1913,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909520.950, "dur": 6.727,
    "args": {
      "External id": 1914,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909529.541, "dur": 5.782,
    "args": {
      "External id": 1915,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909536.533, "dur": 5.614,
    "args": {
      "External id": 1916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909543.645, "dur": 7.163,
    "args": {
      "External id": 1917,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909553.007, "dur": 7.333,
    "args": {
      "External id": 1918,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909553.523, "dur": 0.100,
    "args": {
      "External id": 1919,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909561.831, "dur": 6.343,
    "args": {
      "External id": 1920,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909569.165, "dur": 5.655,
    "args": {
      "External id": 1921,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909598.237, "dur": 6.513,
    "args": {
      "External id": 1922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909599.072, "dur": 5.501,
    "args": {
      "External id": 1923,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909608.142, "dur": 8.393,
    "args": {
      "External id": 1924,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909608.999, "dur": 1.270,
    "args": {
      "External id": 1925,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909610.694, "dur": 5.690,
    "args": {
      "External id": 1926,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909636.827, "dur": 0.162,
    "args": {
      "External id": 1927,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909642.054, "dur": 75.184,
    "args": {
      "External id": 1928,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909648.387, "dur": 1.963,
    "args": {
      "External id": 1929,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909649.488, "dur": 0.399,
    "args": {
      "External id": 1930,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909651.566, "dur": 1.272,
    "args": {
      "External id": 1931,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909652.283, "dur": 0.253,
    "args": {
      "External id": 1932,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909654.775, "dur": 5.924,
    "args": {
      "External id": 1933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909662.634, "dur": 6.905,
    "args": {
      "External id": 1934,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909671.731, "dur": 5.794,
    "args": {
      "External id": 1935,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909678.756, "dur": 5.430,
    "args": {
      "External id": 1936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909685.741, "dur": 6.187,
    "args": {
      "External id": 1937,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909694.058, "dur": 5.726,
    "args": {
      "External id": 1938,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909694.591, "dur": 0.092,
    "args": {
      "External id": 1939,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909701.280, "dur": 5.884,
    "args": {
      "External id": 1940,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909708.138, "dur": 6.107,
    "args": {
      "External id": 1941,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909736.194, "dur": 6.265,
    "args": {
      "External id": 1942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909736.990, "dur": 5.201,
    "args": {
      "External id": 1943,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909745.993, "dur": 8.086,
    "args": {
      "External id": 1944,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909746.780, "dur": 1.285,
    "args": {
      "External id": 1945,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909748.512, "dur": 5.438,
    "args": {
      "External id": 1946,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909773.867, "dur": 0.109,
    "args": {
      "External id": 1947,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909779.017, "dur": 76.802,
    "args": {
      "External id": 1948,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909785.396, "dur": 1.975,
    "args": {
      "External id": 1949,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909786.496, "dur": 0.406,
    "args": {
      "External id": 1950,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909788.493, "dur": 1.309,
    "args": {
      "External id": 1951,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909789.243, "dur": 0.259,
    "args": {
      "External id": 1952,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909791.715, "dur": 6.210,
    "args": {
      "External id": 1953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909799.834, "dur": 7.068,
    "args": {
      "External id": 1954,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909808.905, "dur": 6.007,
    "args": {
      "External id": 1955,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909816.228, "dur": 5.779,
    "args": {
      "External id": 1956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909823.606, "dur": 6.829,
    "args": {
      "External id": 1957,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909832.497, "dur": 5.789,
    "args": {
      "External id": 1958,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909833.059, "dur": 0.094,
    "args": {
      "External id": 1959,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909839.794, "dur": 6.131,
    "args": {
      "External id": 1960,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909846.931, "dur": 5.918,
    "args": {
      "External id": 1961,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740909875.209, "dur": 6.043,
    "args": {
      "External id": 1962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740909876.004, "dur": 5.015,
    "args": {
      "External id": 1963,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740909884.554, "dur": 8.496,
    "args": {
      "External id": 1964,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909885.351, "dur": 1.337,
    "args": {
      "External id": 1965,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740909887.146, "dur": 5.753,
    "args": {
      "External id": 1966,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909913.310, "dur": 0.153,
    "args": {
      "External id": 1967,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909918.314, "dur": 77.706,
    "args": {
      "External id": 1968,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909924.780, "dur": 1.958,
    "args": {
      "External id": 1969,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909925.867, "dur": 0.391,
    "args": {
      "External id": 1970,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740909927.937, "dur": 1.309,
    "args": {
      "External id": 1971,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740909928.670, "dur": 0.275,
    "args": {
      "External id": 1972,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740909931.097, "dur": 5.954,
    "args": {
      "External id": 1973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909938.977, "dur": 7.212,
    "args": {
      "External id": 1974,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909948.134, "dur": 6.153,
    "args": {
      "External id": 1975,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740909955.603, "dur": 5.881,
    "args": {
      "External id": 1976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740909963.059, "dur": 6.832,
    "args": {
      "External id": 1977,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740909971.964, "dur": 6.714,
    "args": {
      "External id": 1978,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740909972.551, "dur": 0.086,
    "args": {
      "External id": 1979,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740909980.151, "dur": 6.169,
    "args": {
      "External id": 1980,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740909987.271, "dur": 5.751,
    "args": {
      "External id": 1981,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910016.276, "dur": 6.280,
    "args": {
      "External id": 1982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910017.049, "dur": 5.278,
    "args": {
      "External id": 1983,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 1982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910025.826, "dur": 10.165,
    "args": {
      "External id": 1984,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 1983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910026.671, "dur": 2.065,
    "args": {
      "External id": 1985,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910029.175, "dur": 6.668,
    "args": {
      "External id": 1986,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910055.649, "dur": 0.157,
    "args": {
      "External id": 1987,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 1986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910061.501, "dur": 121.728,
    "args": {
      "External id": 1988,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 1987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910067.819, "dur": 16.955,
    "args": {
      "External id": 1989,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910082.661, "dur": 1.033,
    "args": {
      "External id": 1990,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910087.401, "dur": 2.038,
    "args": {
      "External id": 1991,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910088.841, "dur": 0.292,
    "args": {
      "External id": 1992,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 1991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910092.239, "dur": 10.190,
    "args": {
      "External id": 1993,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 1992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910105.223, "dur": 8.287,
    "args": {
      "External id": 1994,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910116.035, "dur": 6.290,
    "args": {
      "External id": 1995,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 1994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910123.761, "dur": 5.871,
    "args": {
      "External id": 1996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910131.421, "dur": 10.854,
    "args": {
      "External id": 1997,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910144.607, "dur": 8.560,
    "args": {
      "External id": 1998,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 1997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910145.276, "dur": 0.106,
    "args": {
      "External id": 1999,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 1998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910154.869, "dur": 17.053,
    "args": {
      "External id": 2000,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 1999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910173.091, "dur": 6.387,
    "args": {
      "External id": 2001,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910206.604, "dur": 7.644,
    "args": {
      "External id": 2002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910207.759, "dur": 6.206,
    "args": {
      "External id": 2003,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910218.136, "dur": 9.659,
    "args": {
      "External id": 2004,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910219.344, "dur": 1.584,
    "args": {
      "External id": 2005,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910221.426, "dur": 6.211,
    "args": {
      "External id": 2006,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910249.901, "dur": 0.151,
    "args": {
      "External id": 2007,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910256.062, "dur": 81.791,
    "args": {
      "External id": 2008,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910263.865, "dur": 2.036,
    "args": {
      "External id": 2009,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910264.998, "dur": 0.435,
    "args": {
      "External id": 2010,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910267.222, "dur": 1.624,
    "args": {
      "External id": 2011,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910268.101, "dur": 0.461,
    "args": {
      "External id": 2012,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910270.922, "dur": 6.752,
    "args": {
      "External id": 2013,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910280.002, "dur": 6.797,
    "args": {
      "External id": 2014,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910288.935, "dur": 6.356,
    "args": {
      "External id": 2015,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910296.703, "dur": 6.279,
    "args": {
      "External id": 2016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910304.658, "dur": 6.691,
    "args": {
      "External id": 2017,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910313.599, "dur": 6.672,
    "args": {
      "External id": 2018,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910314.225, "dur": 0.088,
    "args": {
      "External id": 2019,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910321.728, "dur": 6.128,
    "args": {
      "External id": 2020,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910328.791, "dur": 5.693,
    "args": {
      "External id": 2021,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910359.129, "dur": 6.303,
    "args": {
      "External id": 2022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910359.975, "dur": 5.234,
    "args": {
      "External id": 2023,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910368.690, "dur": 8.990,
    "args": {
      "External id": 2024,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910369.586, "dur": 1.386,
    "args": {
      "External id": 2025,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910371.479, "dur": 6.065,
    "args": {
      "External id": 2026,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910399.364, "dur": 0.133,
    "args": {
      "External id": 2027,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910404.721, "dur": 80.923,
    "args": {
      "External id": 2028,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910411.746, "dur": 2.115,
    "args": {
      "External id": 2029,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910412.976, "dur": 0.343,
    "args": {
      "External id": 2030,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910415.051, "dur": 1.313,
    "args": {
      "External id": 2031,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910415.814, "dur": 0.243,
    "args": {
      "External id": 2032,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910418.341, "dur": 7.301,
    "args": {
      "External id": 2033,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910427.785, "dur": 6.559,
    "args": {
      "External id": 2034,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910436.408, "dur": 7.499,
    "args": {
      "External id": 2035,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910445.376, "dur": 6.245,
    "args": {
      "External id": 2036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910453.278, "dur": 6.798,
    "args": {
      "External id": 2037,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910462.277, "dur": 6.252,
    "args": {
      "External id": 2038,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910462.880, "dur": 0.076,
    "args": {
      "External id": 2039,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910469.936, "dur": 6.251,
    "args": {
      "External id": 2040,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910477.121, "dur": 5.469,
    "args": {
      "External id": 2041,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910506.308, "dur": 5.760,
    "args": {
      "External id": 2042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910507.222, "dur": 4.620,
    "args": {
      "External id": 2043,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910515.435, "dur": 8.691,
    "args": {
      "External id": 2044,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910516.286, "dur": 1.257,
    "args": {
      "External id": 2045,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910518.086, "dur": 5.877,
    "args": {
      "External id": 2046,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910544.963, "dur": 0.265,
    "args": {
      "External id": 2047,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910550.853, "dur": 101.640,
    "args": {
      "External id": 2048,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910566.759, "dur": 2.275,
    "args": {
      "External id": 2049,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910568.089, "dur": 0.403,
    "args": {
      "External id": 2050,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910570.380, "dur": 1.452,
    "args": {
      "External id": 2051,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910571.256, "dur": 0.273,
    "args": {
      "External id": 2052,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910574.090, "dur": 7.141,
    "args": {
      "External id": 2053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910583.492, "dur": 7.420,
    "args": {
      "External id": 2054,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910593.007, "dur": 5.904,
    "args": {
      "External id": 2055,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910600.380, "dur": 10.570,
    "args": {
      "External id": 2056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910612.618, "dur": 10.869,
    "args": {
      "External id": 2057,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910625.986, "dur": 7.433,
    "args": {
      "External id": 2058,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910626.705, "dur": 0.100,
    "args": {
      "External id": 2059,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910635.057, "dur": 6.509,
    "args": {
      "External id": 2060,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910642.647, "dur": 6.456,
    "args": {
      "External id": 2061,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910674.780, "dur": 6.962,
    "args": {
      "External id": 2062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910675.693, "dur": 5.745,
    "args": {
      "External id": 2063,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910685.520, "dur": 9.298,
    "args": {
      "External id": 2064,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910686.496, "dur": 1.350,
    "args": {
      "External id": 2065,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910688.421, "dur": 6.245,
    "args": {
      "External id": 2066,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910717.846, "dur": 0.133,
    "args": {
      "External id": 2067,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910723.645, "dur": 86.210,
    "args": {
      "External id": 2068,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910731.395, "dur": 2.576,
    "args": {
      "External id": 2069,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910733.041, "dur": 0.397,
    "args": {
      "External id": 2070,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910735.362, "dur": 1.430,
    "args": {
      "External id": 2071,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910736.167, "dur": 0.291,
    "args": {
      "External id": 2072,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910738.974, "dur": 7.307,
    "args": {
      "External id": 2073,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910748.490, "dur": 6.780,
    "args": {
      "External id": 2074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910757.529, "dur": 7.709,
    "args": {
      "External id": 2075,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910766.823, "dur": 5.805,
    "args": {
      "External id": 2076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910774.388, "dur": 7.150,
    "args": {
      "External id": 2077,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910783.755, "dur": 6.790,
    "args": {
      "External id": 2078,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910784.422, "dur": 0.089,
    "args": {
      "External id": 2079,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910792.138, "dur": 7.118,
    "args": {
      "External id": 2080,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910800.311, "dur": 6.245,
    "args": {
      "External id": 2081,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910831.357, "dur": 7.653,
    "args": {
      "External id": 2082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910832.211, "dur": 6.544,
    "args": {
      "External id": 2083,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910842.688, "dur": 9.616,
    "args": {
      "External id": 2084,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910843.690, "dur": 1.473,
    "args": {
      "External id": 2085,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910845.640, "dur": 6.473,
    "args": {
      "External id": 2086,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910875.539, "dur": 0.155,
    "args": {
      "External id": 2087,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910881.028, "dur": 82.250,
    "args": {
      "External id": 2088,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910888.306, "dur": 2.194,
    "args": {
      "External id": 2089,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910889.505, "dur": 0.451,
    "args": {
      "External id": 2090,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740910891.811, "dur": 1.427,
    "args": {
      "External id": 2091,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910892.637, "dur": 0.283,
    "args": {
      "External id": 2092,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740910895.451, "dur": 6.165,
    "args": {
      "External id": 2093,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910903.733, "dur": 7.434,
    "args": {
      "External id": 2094,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910913.913, "dur": 6.247,
    "args": {
      "External id": 2095,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740910921.632, "dur": 6.022,
    "args": {
      "External id": 2096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740910929.384, "dur": 6.700,
    "args": {
      "External id": 2097,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740910938.397, "dur": 6.832,
    "args": {
      "External id": 2098,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740910939.081, "dur": 0.092,
    "args": {
      "External id": 2099,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740910946.837, "dur": 6.468,
    "args": {
      "External id": 2100,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740910954.337, "dur": 5.678,
    "args": {
      "External id": 2101,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740910985.029, "dur": 6.808,
    "args": {
      "External id": 2102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740910985.959, "dur": 5.632,
    "args": {
      "External id": 2103,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740910995.459, "dur": 8.832,
    "args": {
      "External id": 2104,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740910996.386, "dur": 1.417,
    "args": {
      "External id": 2105,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740910998.288, "dur": 5.827,
    "args": {
      "External id": 2106,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911026.793, "dur": 0.137,
    "args": {
      "External id": 2107,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911032.267, "dur": 80.450,
    "args": {
      "External id": 2108,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911039.348, "dur": 2.038,
    "args": {
      "External id": 2109,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911040.503, "dur": 0.404,
    "args": {
      "External id": 2110,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911042.679, "dur": 1.459,
    "args": {
      "External id": 2111,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911043.516, "dur": 0.290,
    "args": {
      "External id": 2112,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911046.304, "dur": 6.453,
    "args": {
      "External id": 2113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911054.977, "dur": 7.248,
    "args": {
      "External id": 2114,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911064.619, "dur": 6.083,
    "args": {
      "External id": 2115,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911072.201, "dur": 5.911,
    "args": {
      "External id": 2116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911079.764, "dur": 6.578,
    "args": {
      "External id": 2117,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911088.500, "dur": 6.688,
    "args": {
      "External id": 2118,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911089.138, "dur": 0.091,
    "args": {
      "External id": 2119,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911096.735, "dur": 5.741,
    "args": {
      "External id": 2120,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911103.488, "dur": 5.967,
    "args": {
      "External id": 2121,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911133.921, "dur": 6.433,
    "args": {
      "External id": 2122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911134.758, "dur": 5.311,
    "args": {
      "External id": 2123,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911143.977, "dur": 8.614,
    "args": {
      "External id": 2124,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911144.882, "dur": 1.353,
    "args": {
      "External id": 2125,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911146.755, "dur": 5.668,
    "args": {
      "External id": 2126,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911174.487, "dur": 0.119,
    "args": {
      "External id": 2127,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911179.819, "dur": 104.450,
    "args": {
      "External id": 2128,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911187.968, "dur": 16.337,
    "args": {
      "External id": 2129,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911203.205, "dur": 0.543,
    "args": {
      "External id": 2130,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911205.822, "dur": 1.690,
    "args": {
      "External id": 2131,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911206.837, "dur": 0.325,
    "args": {
      "External id": 2132,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911209.840, "dur": 7.322,
    "args": {
      "External id": 2133,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911219.704, "dur": 7.515,
    "args": {
      "External id": 2134,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911229.658, "dur": 7.874,
    "args": {
      "External id": 2135,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911239.090, "dur": 6.738,
    "args": {
      "External id": 2136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911247.650, "dur": 7.812,
    "args": {
      "External id": 2137,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911257.942, "dur": 7.166,
    "args": {
      "External id": 2138,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911258.651, "dur": 0.121,
    "args": {
      "External id": 2139,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911266.713, "dur": 6.673,
    "args": {
      "External id": 2140,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911274.477, "dur": 6.003,
    "args": {
      "External id": 2141,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911308.023, "dur": 6.123,
    "args": {
      "External id": 2142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911308.899, "dur": 4.991,
    "args": {
      "External id": 2143,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911317.859, "dur": 9.373,
    "args": {
      "External id": 2144,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911319.502, "dur": 1.461,
    "args": {
      "External id": 2145,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911321.484, "dur": 5.587,
    "args": {
      "External id": 2146,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911351.018, "dur": 0.164,
    "args": {
      "External id": 2147,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911357.402, "dur": 84.898,
    "args": {
      "External id": 2148,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911365.847, "dur": 2.160,
    "args": {
      "External id": 2149,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911367.128, "dur": 0.408,
    "args": {
      "External id": 2150,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911369.358, "dur": 1.720,
    "args": {
      "External id": 2151,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911370.493, "dur": 0.271,
    "args": {
      "External id": 2152,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911373.288, "dur": 6.110,
    "args": {
      "External id": 2153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911381.652, "dur": 7.465,
    "args": {
      "External id": 2154,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911391.554, "dur": 6.838,
    "args": {
      "External id": 2155,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911399.838, "dur": 6.182,
    "args": {
      "External id": 2156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911407.791, "dur": 5.835,
    "args": {
      "External id": 2157,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911415.994, "dur": 7.156,
    "args": {
      "External id": 2158,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911416.623, "dur": 0.093,
    "args": {
      "External id": 2159,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911424.717, "dur": 6.762,
    "args": {
      "External id": 2160,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911432.533, "dur": 6.289,
    "args": {
      "External id": 2161,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911463.628, "dur": 6.669,
    "args": {
      "External id": 2162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911464.556, "dur": 5.493,
    "args": {
      "External id": 2163,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911474.032, "dur": 9.492,
    "args": {
      "External id": 2164,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911474.940, "dur": 1.500,
    "args": {
      "External id": 2165,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911477.044, "dur": 6.322,
    "args": {
      "External id": 2166,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911506.102, "dur": 0.182,
    "args": {
      "External id": 2167,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911511.974, "dur": 81.827,
    "args": {
      "External id": 2168,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911519.717, "dur": 2.197,
    "args": {
      "External id": 2169,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911520.993, "dur": 0.406,
    "args": {
      "External id": 2170,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911523.193, "dur": 1.412,
    "args": {
      "External id": 2171,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911523.993, "dur": 0.291,
    "args": {
      "External id": 2172,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911526.683, "dur": 6.656,
    "args": {
      "External id": 2173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911535.428, "dur": 6.860,
    "args": {
      "External id": 2174,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911544.498, "dur": 5.579,
    "args": {
      "External id": 2175,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911551.535, "dur": 6.079,
    "args": {
      "External id": 2176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911559.257, "dur": 6.940,
    "args": {
      "External id": 2177,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911568.611, "dur": 7.094,
    "args": {
      "External id": 2178,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911569.305, "dur": 0.089,
    "args": {
      "External id": 2179,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911577.227, "dur": 6.324,
    "args": {
      "External id": 2180,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911584.597, "dur": 5.914,
    "args": {
      "External id": 2181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911620.282, "dur": 7.192,
    "args": {
      "External id": 2182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911621.190, "dur": 6.067,
    "args": {
      "External id": 2183,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911631.184, "dur": 9.327,
    "args": {
      "External id": 2184,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911632.159, "dur": 1.485,
    "args": {
      "External id": 2185,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911634.145, "dur": 6.196,
    "args": {
      "External id": 2186,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911662.846, "dur": 0.149,
    "args": {
      "External id": 2187,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911668.486, "dur": 86.779,
    "args": {
      "External id": 2188,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911675.916, "dur": 2.227,
    "args": {
      "External id": 2189,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911677.213, "dur": 0.420,
    "args": {
      "External id": 2190,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911679.388, "dur": 1.446,
    "args": {
      "External id": 2191,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911680.235, "dur": 0.285,
    "args": {
      "External id": 2192,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911682.862, "dur": 7.478,
    "args": {
      "External id": 2193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911692.478, "dur": 7.872,
    "args": {
      "External id": 2194,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911702.611, "dur": 6.796,
    "args": {
      "External id": 2195,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911710.915, "dur": 6.352,
    "args": {
      "External id": 2196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911718.852, "dur": 6.947,
    "args": {
      "External id": 2197,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911728.126, "dur": 7.149,
    "args": {
      "External id": 2198,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911728.880, "dur": 0.090,
    "args": {
      "External id": 2199,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911736.853, "dur": 7.326,
    "args": {
      "External id": 2200,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911745.325, "dur": 6.635,
    "args": {
      "External id": 2201,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911777.492, "dur": 7.371,
    "args": {
      "External id": 2202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911778.307, "dur": 6.290,
    "args": {
      "External id": 2203,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911788.498, "dur": 11.497,
    "args": {
      "External id": 2204,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911790.482, "dur": 1.418,
    "args": {
      "External id": 2205,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911792.347, "dur": 7.499,
    "args": {
      "External id": 2206,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911822.236, "dur": 0.163,
    "args": {
      "External id": 2207,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911829.005, "dur": 88.442,
    "args": {
      "External id": 2208,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911837.639, "dur": 2.279,
    "args": {
      "External id": 2209,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911838.996, "dur": 0.393,
    "args": {
      "External id": 2210,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911841.240, "dur": 1.600,
    "args": {
      "External id": 2211,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911842.234, "dur": 0.295,
    "args": {
      "External id": 2212,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740911845.010, "dur": 8.088,
    "args": {
      "External id": 2213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911855.368, "dur": 7.409,
    "args": {
      "External id": 2214,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911865.077, "dur": 6.650,
    "args": {
      "External id": 2215,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740911873.066, "dur": 6.242,
    "args": {
      "External id": 2216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740911881.084, "dur": 7.176,
    "args": {
      "External id": 2217,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740911890.655, "dur": 6.978,
    "args": {
      "External id": 2218,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911891.413, "dur": 0.086,
    "args": {
      "External id": 2219,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740911899.145, "dur": 6.848,
    "args": {
      "External id": 2220,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740911907.008, "dur": 7.014,
    "args": {
      "External id": 2221,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740911939.089, "dur": 7.480,
    "args": {
      "External id": 2222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740911939.996, "dur": 6.356,
    "args": {
      "External id": 2223,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740911950.127, "dur": 9.250,
    "args": {
      "External id": 2224,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911950.969, "dur": 1.454,
    "args": {
      "External id": 2225,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740911953.039, "dur": 6.143,
    "args": {
      "External id": 2226,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740911981.056, "dur": 0.145,
    "args": {
      "External id": 2227,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911986.674, "dur": 81.489,
    "args": {
      "External id": 2228,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911994.299, "dur": 2.109,
    "args": {
      "External id": 2229,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911995.484, "dur": 0.422,
    "args": {
      "External id": 2230,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740911997.670, "dur": 1.449,
    "args": {
      "External id": 2231,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740911998.507, "dur": 0.280,
    "args": {
      "External id": 2232,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912001.192, "dur": 5.942,
    "args": {
      "External id": 2233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912009.292, "dur": 7.176,
    "args": {
      "External id": 2234,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912018.712, "dur": 6.409,
    "args": {
      "External id": 2235,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912026.560, "dur": 6.192,
    "args": {
      "External id": 2236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912034.392, "dur": 6.299,
    "args": {
      "External id": 2237,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912043.032, "dur": 6.269,
    "args": {
      "External id": 2238,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912043.667, "dur": 0.089,
    "args": {
      "External id": 2239,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912050.801, "dur": 6.723,
    "args": {
      "External id": 2240,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912058.537, "dur": 6.440,
    "args": {
      "External id": 2241,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912088.752, "dur": 6.288,
    "args": {
      "External id": 2242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912089.598, "dur": 5.154,
    "args": {
      "External id": 2243,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912098.574, "dur": 9.172,
    "args": {
      "External id": 2244,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912099.486, "dur": 1.368,
    "args": {
      "External id": 2245,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912101.375, "dur": 6.186,
    "args": {
      "External id": 2246,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912129.198, "dur": 0.154,
    "args": {
      "External id": 2247,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912134.604, "dur": 82.919,
    "args": {
      "External id": 2248,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912141.701, "dur": 2.097,
    "args": {
      "External id": 2249,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912142.921, "dur": 0.389,
    "args": {
      "External id": 2250,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912145.084, "dur": 1.439,
    "args": {
      "External id": 2251,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912145.911, "dur": 0.284,
    "args": {
      "External id": 2252,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912150.429, "dur": 7.235,
    "args": {
      "External id": 2253,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912160.017, "dur": 6.694,
    "args": {
      "External id": 2254,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912168.820, "dur": 6.073,
    "args": {
      "External id": 2255,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912176.259, "dur": 5.953,
    "args": {
      "External id": 2256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912183.872, "dur": 5.878,
    "args": {
      "External id": 2257,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912192.046, "dur": 6.966,
    "args": {
      "External id": 2258,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912192.663, "dur": 0.100,
    "args": {
      "External id": 2259,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912200.517, "dur": 6.087,
    "args": {
      "External id": 2260,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912207.650, "dur": 6.498,
    "args": {
      "External id": 2261,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912237.712, "dur": 6.567,
    "args": {
      "External id": 2262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912238.606, "dur": 5.437,
    "args": {
      "External id": 2263,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912247.631, "dur": 8.668,
    "args": {
      "External id": 2264,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912248.545, "dur": 1.458,
    "args": {
      "External id": 2265,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912250.452, "dur": 5.629,
    "args": {
      "External id": 2266,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912276.996, "dur": 0.131,
    "args": {
      "External id": 2267,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912282.477, "dur": 99.102,
    "args": {
      "External id": 2268,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912289.679, "dur": 2.057,
    "args": {
      "External id": 2269,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912290.804, "dur": 0.443,
    "args": {
      "External id": 2270,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912293.013, "dur": 1.433,
    "args": {
      "External id": 2271,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912293.851, "dur": 0.283,
    "args": {
      "External id": 2272,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912296.549, "dur": 7.181,
    "args": {
      "External id": 2273,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912306.126, "dur": 6.860,
    "args": {
      "External id": 2274,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912315.071, "dur": 6.147,
    "args": {
      "External id": 2275,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912322.659, "dur": 5.824,
    "args": {
      "External id": 2276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912330.106, "dur": 6.172,
    "args": {
      "External id": 2277,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912338.512, "dur": 7.130,
    "args": {
      "External id": 2278,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912339.201, "dur": 0.091,
    "args": {
      "External id": 2279,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912347.121, "dur": 16.219,
    "args": {
      "External id": 2280,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912365.881, "dur": 10.836,
    "args": {
      "External id": 2281,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912406.825, "dur": 8.023,
    "args": {
      "External id": 2282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912407.938, "dur": 6.617,
    "args": {
      "External id": 2283,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912418.952, "dur": 10.316,
    "args": {
      "External id": 2284,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912420.217, "dur": 1.524,
    "args": {
      "External id": 2285,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912422.296, "dur": 6.785,
    "args": {
      "External id": 2286,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912452.797, "dur": 0.176,
    "args": {
      "External id": 2287,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912459.282, "dur": 92.889,
    "args": {
      "External id": 2288,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912467.881, "dur": 2.652,
    "args": {
      "External id": 2289,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912469.386, "dur": 0.497,
    "args": {
      "External id": 2290,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912471.890, "dur": 2.259,
    "args": {
      "External id": 2291,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912473.058, "dur": 0.785,
    "args": {
      "External id": 2292,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912476.424, "dur": 7.309,
    "args": {
      "External id": 2293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912486.414, "dur": 7.179,
    "args": {
      "External id": 2294,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912496.057, "dur": 7.866,
    "args": {
      "External id": 2295,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912505.400, "dur": 5.990,
    "args": {
      "External id": 2296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912513.097, "dur": 8.845,
    "args": {
      "External id": 2297,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912524.394, "dur": 7.829,
    "args": {
      "External id": 2298,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912525.193, "dur": 0.088,
    "args": {
      "External id": 2299,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912533.894, "dur": 7.123,
    "args": {
      "External id": 2300,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912542.035, "dur": 6.474,
    "args": {
      "External id": 2301,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912574.145, "dur": 7.864,
    "args": {
      "External id": 2302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912575.062, "dur": 6.736,
    "args": {
      "External id": 2303,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912585.641, "dur": 9.657,
    "args": {
      "External id": 2304,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912586.567, "dur": 1.474,
    "args": {
      "External id": 2305,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912588.558, "dur": 6.593,
    "args": {
      "External id": 2306,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912618.513, "dur": 0.139,
    "args": {
      "External id": 2307,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912625.089, "dur": 86.176,
    "args": {
      "External id": 2308,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912633.739, "dur": 2.073,
    "args": {
      "External id": 2309,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912634.879, "dur": 0.403,
    "args": {
      "External id": 2310,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912637.124, "dur": 1.451,
    "args": {
      "External id": 2311,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912637.970, "dur": 0.281,
    "args": {
      "External id": 2312,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912641.078, "dur": 6.438,
    "args": {
      "External id": 2313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912649.695, "dur": 7.748,
    "args": {
      "External id": 2314,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912660.190, "dur": 6.468,
    "args": {
      "External id": 2315,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912668.134, "dur": 5.985,
    "args": {
      "External id": 2316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912675.911, "dur": 6.757,
    "args": {
      "External id": 2317,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912684.957, "dur": 7.293,
    "args": {
      "External id": 2318,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912685.588, "dur": 0.102,
    "args": {
      "External id": 2319,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912693.908, "dur": 6.764,
    "args": {
      "External id": 2320,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912701.765, "dur": 5.872,
    "args": {
      "External id": 2321,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912733.547, "dur": 7.834,
    "args": {
      "External id": 2322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912734.419, "dur": 6.732,
    "args": {
      "External id": 2323,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912744.906, "dur": 8.990,
    "args": {
      "External id": 2324,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912745.844, "dur": 1.316,
    "args": {
      "External id": 2325,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912747.650, "dur": 6.093,
    "args": {
      "External id": 2326,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912776.276, "dur": 0.139,
    "args": {
      "External id": 2327,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912782.246, "dur": 82.346,
    "args": {
      "External id": 2328,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912789.787, "dur": 2.127,
    "args": {
      "External id": 2329,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912790.968, "dur": 0.404,
    "args": {
      "External id": 2330,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912793.185, "dur": 1.440,
    "args": {
      "External id": 2331,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912794.007, "dur": 0.294,
    "args": {
      "External id": 2332,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912797.097, "dur": 6.892,
    "args": {
      "External id": 2333,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912806.363, "dur": 6.692,
    "args": {
      "External id": 2334,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912815.149, "dur": 6.425,
    "args": {
      "External id": 2335,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912823.022, "dur": 5.900,
    "args": {
      "External id": 2336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912830.538, "dur": 5.993,
    "args": {
      "External id": 2337,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912838.797, "dur": 6.996,
    "args": {
      "External id": 2338,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912839.476, "dur": 0.099,
    "args": {
      "External id": 2339,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740912847.398, "dur": 6.576,
    "args": {
      "External id": 2340,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912855.000, "dur": 6.057,
    "args": {
      "External id": 2341,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740912886.429, "dur": 6.546,
    "args": {
      "External id": 2342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740912887.392, "dur": 5.370,
    "args": {
      "External id": 2343,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740912896.531, "dur": 8.894,
    "args": {
      "External id": 2344,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912897.571, "dur": 1.393,
    "args": {
      "External id": 2345,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740912899.501, "dur": 5.785,
    "args": {
      "External id": 2346,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912930.145, "dur": 0.147,
    "args": {
      "External id": 2347,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912938.336, "dur": 86.328,
    "args": {
      "External id": 2348,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912945.892, "dur": 2.617,
    "args": {
      "External id": 2349,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912947.543, "dur": 0.431,
    "args": {
      "External id": 2350,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740912949.828, "dur": 1.807,
    "args": {
      "External id": 2351,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740912950.830, "dur": 0.473,
    "args": {
      "External id": 2352,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740912953.823, "dur": 7.312,
    "args": {
      "External id": 2353,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912963.547, "dur": 8.184,
    "args": {
      "External id": 2354,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740912973.947, "dur": 6.685,
    "args": {
      "External id": 2355,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740912982.070, "dur": 6.554,
    "args": {
      "External id": 2356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740912990.335, "dur": 6.713,
    "args": {
      "External id": 2357,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740912999.284, "dur": 6.919,
    "args": {
      "External id": 2358,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740912999.934, "dur": 0.103,
    "args": {
      "External id": 2359,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913007.799, "dur": 6.224,
    "args": {
      "External id": 2360,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913015.069, "dur": 6.169,
    "args": {
      "External id": 2361,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913046.258, "dur": 5.872,
    "args": {
      "External id": 2362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913047.138, "dur": 4.763,
    "args": {
      "External id": 2363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913055.787, "dur": 9.969,
    "args": {
      "External id": 2364,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913057.565, "dur": 1.408,
    "args": {
      "External id": 2365,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913059.468, "dur": 6.140,
    "args": {
      "External id": 2366,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913088.413, "dur": 0.129,
    "args": {
      "External id": 2367,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913093.940, "dur": 81.101,
    "args": {
      "External id": 2368,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913101.071, "dur": 2.130,
    "args": {
      "External id": 2369,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913102.293, "dur": 0.395,
    "args": {
      "External id": 2370,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913104.523, "dur": 1.429,
    "args": {
      "External id": 2371,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913105.367, "dur": 0.270,
    "args": {
      "External id": 2372,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913108.244, "dur": 6.247,
    "args": {
      "External id": 2373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913116.675, "dur": 7.030,
    "args": {
      "External id": 2374,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913125.930, "dur": 6.105,
    "args": {
      "External id": 2375,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913133.454, "dur": 5.705,
    "args": {
      "External id": 2376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913140.794, "dur": 6.492,
    "args": {
      "External id": 2377,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913149.621, "dur": 7.302,
    "args": {
      "External id": 2378,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913150.253, "dur": 0.094,
    "args": {
      "External id": 2379,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913158.590, "dur": 6.192,
    "args": {
      "External id": 2380,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913165.831, "dur": 5.970,
    "args": {
      "External id": 2381,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913195.866, "dur": 7.104,
    "args": {
      "External id": 2382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913196.804, "dur": 5.925,
    "args": {
      "External id": 2383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913206.588, "dur": 8.309,
    "args": {
      "External id": 2384,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913207.531, "dur": 1.408,
    "args": {
      "External id": 2385,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913209.465, "dur": 5.280,
    "args": {
      "External id": 2386,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913237.264, "dur": 0.126,
    "args": {
      "External id": 2387,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913242.884, "dur": 82.605,
    "args": {
      "External id": 2388,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913249.961, "dur": 2.161,
    "args": {
      "External id": 2389,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913251.180, "dur": 0.434,
    "args": {
      "External id": 2390,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913253.424, "dur": 1.423,
    "args": {
      "External id": 2391,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913254.245, "dur": 0.271,
    "args": {
      "External id": 2392,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913257.009, "dur": 6.503,
    "args": {
      "External id": 2393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913265.658, "dur": 7.043,
    "args": {
      "External id": 2394,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913274.840, "dur": 5.547,
    "args": {
      "External id": 2395,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913281.810, "dur": 6.026,
    "args": {
      "External id": 2396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913289.475, "dur": 7.438,
    "args": {
      "External id": 2397,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913299.606, "dur": 6.540,
    "args": {
      "External id": 2398,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913300.202, "dur": 0.091,
    "args": {
      "External id": 2399,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913307.724, "dur": 7.140,
    "args": {
      "External id": 2400,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913315.901, "dur": 6.230,
    "args": {
      "External id": 2401,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913345.464, "dur": 7.047,
    "args": {
      "External id": 2402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913346.313, "dur": 5.946,
    "args": {
      "External id": 2403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913356.230, "dur": 9.355,
    "args": {
      "External id": 2404,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913357.196, "dur": 1.414,
    "args": {
      "External id": 2405,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913359.098, "dur": 6.327,
    "args": {
      "External id": 2406,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913387.043, "dur": 0.155,
    "args": {
      "External id": 2407,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913392.580, "dur": 82.322,
    "args": {
      "External id": 2408,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913399.772, "dur": 2.040,
    "args": {
      "External id": 2409,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913400.892, "dur": 0.398,
    "args": {
      "External id": 2410,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913403.048, "dur": 1.403,
    "args": {
      "External id": 2411,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913403.832, "dur": 0.305,
    "args": {
      "External id": 2412,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913406.616, "dur": 6.978,
    "args": {
      "External id": 2413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913415.799, "dur": 6.586,
    "args": {
      "External id": 2414,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913424.515, "dur": 6.660,
    "args": {
      "External id": 2415,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913432.558, "dur": 6.054,
    "args": {
      "External id": 2416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913440.300, "dur": 6.496,
    "args": {
      "External id": 2417,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913448.999, "dur": 6.657,
    "args": {
      "External id": 2418,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913449.617, "dur": 0.089,
    "args": {
      "External id": 2419,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913457.218, "dur": 6.453,
    "args": {
      "External id": 2420,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913464.856, "dur": 6.696,
    "args": {
      "External id": 2421,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913495.888, "dur": 6.381,
    "args": {
      "External id": 2422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913496.841, "dur": 5.182,
    "args": {
      "External id": 2423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913505.707, "dur": 20.513,
    "args": {
      "External id": 2424,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913510.226, "dur": 1.453,
    "args": {
      "External id": 2425,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913512.163, "dur": 13.637,
    "args": {
      "External id": 2426,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913555.681, "dur": 0.302,
    "args": {
      "External id": 2427,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913562.887, "dur": 108.968,
    "args": {
      "External id": 2428,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913571.391, "dur": 2.916,
    "args": {
      "External id": 2429,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913572.959, "dur": 0.711,
    "args": {
      "External id": 2430,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913575.588, "dur": 1.606,
    "args": {
      "External id": 2431,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913576.510, "dur": 0.340,
    "args": {
      "External id": 2432,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913579.544, "dur": 23.075,
    "args": {
      "External id": 2433,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913605.705, "dur": 9.060,
    "args": {
      "External id": 2434,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913617.306, "dur": 7.198,
    "args": {
      "External id": 2435,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913626.045, "dur": 6.343,
    "args": {
      "External id": 2436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913634.276, "dur": 7.568,
    "args": {
      "External id": 2437,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913644.393, "dur": 8.482,
    "args": {
      "External id": 2438,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913645.102, "dur": 0.125,
    "args": {
      "External id": 2439,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913654.701, "dur": 5.912,
    "args": {
      "External id": 2440,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913661.724, "dur": 6.284,
    "args": {
      "External id": 2441,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913695.560, "dur": 7.501,
    "args": {
      "External id": 2442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913696.532, "dur": 6.202,
    "args": {
      "External id": 2443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913707.107, "dur": 10.411,
    "args": {
      "External id": 2444,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913708.230, "dur": 1.613,
    "args": {
      "External id": 2445,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913710.398, "dur": 6.934,
    "args": {
      "External id": 2446,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913739.795, "dur": 0.163,
    "args": {
      "External id": 2447,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913745.704, "dur": 83.518,
    "args": {
      "External id": 2448,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913753.494, "dur": 2.095,
    "args": {
      "External id": 2449,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913754.735, "dur": 0.374,
    "args": {
      "External id": 2450,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913756.876, "dur": 1.446,
    "args": {
      "External id": 2451,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913757.706, "dur": 0.287,
    "args": {
      "External id": 2452,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913760.506, "dur": 7.175,
    "args": {
      "External id": 2453,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913769.977, "dur": 6.868,
    "args": {
      "External id": 2454,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913779.260, "dur": 6.263,
    "args": {
      "External id": 2455,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913786.870, "dur": 5.441,
    "args": {
      "External id": 2456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913794.002, "dur": 7.554,
    "args": {
      "External id": 2457,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913803.973, "dur": 6.894,
    "args": {
      "External id": 2458,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913804.642, "dur": 0.091,
    "args": {
      "External id": 2459,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913812.464, "dur": 6.301,
    "args": {
      "External id": 2460,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913819.837, "dur": 5.984,
    "args": {
      "External id": 2461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913851.104, "dur": 6.574,
    "args": {
      "External id": 2462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740913852.027, "dur": 5.350,
    "args": {
      "External id": 2463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740913861.462, "dur": 8.225,
    "args": {
      "External id": 2464,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913862.440, "dur": 1.490,
    "args": {
      "External id": 2465,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740913864.482, "dur": 5.048,
    "args": {
      "External id": 2466,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913892.482, "dur": 0.166,
    "args": {
      "External id": 2467,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913898.067, "dur": 80.707,
    "args": {
      "External id": 2468,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913905.402, "dur": 2.049,
    "args": {
      "External id": 2469,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913906.544, "dur": 0.424,
    "args": {
      "External id": 2470,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740913908.773, "dur": 1.425,
    "args": {
      "External id": 2471,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740913909.613, "dur": 0.273,
    "args": {
      "External id": 2472,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740913912.382, "dur": 6.580,
    "args": {
      "External id": 2473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913921.269, "dur": 6.436,
    "args": {
      "External id": 2474,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913929.855, "dur": 5.989,
    "args": {
      "External id": 2475,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740913937.329, "dur": 6.063,
    "args": {
      "External id": 2476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740913945.001, "dur": 6.270,
    "args": {
      "External id": 2477,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740913953.611, "dur": 6.544,
    "args": {
      "External id": 2478,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740913954.237, "dur": 0.129,
    "args": {
      "External id": 2479,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740913961.731, "dur": 6.637,
    "args": {
      "External id": 2480,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740913969.427, "dur": 5.881,
    "args": {
      "External id": 2481,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740913999.784, "dur": 6.150,
    "args": {
      "External id": 2482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914000.631, "dur": 5.025,
    "args": {
      "External id": 2483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914009.691, "dur": 10.074,
    "args": {
      "External id": 2484,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914010.685, "dur": 1.460,
    "args": {
      "External id": 2485,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914012.640, "dur": 6.990,
    "args": {
      "External id": 2486,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914041.521, "dur": 0.159,
    "args": {
      "External id": 2487,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914046.801, "dur": 129.046,
    "args": {
      "External id": 2488,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914053.805, "dur": 2.077,
    "args": {
      "External id": 2489,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914054.977, "dur": 0.424,
    "args": {
      "External id": 2490,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914057.165, "dur": 1.369,
    "args": {
      "External id": 2491,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914057.952, "dur": 0.262,
    "args": {
      "External id": 2492,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914060.737, "dur": 6.727,
    "args": {
      "External id": 2493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914102.371, "dur": 11.212,
    "args": {
      "External id": 2494,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914117.470, "dur": 6.686,
    "args": {
      "External id": 2495,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914125.894, "dur": 7.823,
    "args": {
      "External id": 2496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914135.734, "dur": 8.254,
    "args": {
      "External id": 2497,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914146.766, "dur": 8.701,
    "args": {
      "External id": 2498,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914147.734, "dur": 0.152,
    "args": {
      "External id": 2499,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740914157.280, "dur": 6.770,
    "args": {
      "External id": 2500,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914165.172, "dur": 6.571,
    "args": {
      "External id": 2501,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740914202.353, "dur": 7.392,
    "args": {
      "External id": 2502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914203.456, "dur": 5.959,
    "args": {
      "External id": 2503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914213.820, "dur": 10.848,
    "args": {
      "External id": 2504,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914215.271, "dur": 1.641,
    "args": {
      "External id": 2505,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914217.504, "dur": 7.017,
    "args": {
      "External id": 2506,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914248.404, "dur": 0.146,
    "args": {
      "External id": 2507,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914255.192, "dur": 89.119,
    "args": {
      "External id": 2508,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914263.988, "dur": 2.605,
    "args": {
      "External id": 2509,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914265.360, "dur": 0.592,
    "args": {
      "External id": 2510,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914267.976, "dur": 1.468,
    "args": {
      "External id": 2511,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914268.821, "dur": 0.285,
    "args": {
      "External id": 2512,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914271.872, "dur": 9.445,
    "args": {
      "External id": 2513,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914284.172, "dur": 6.776,
    "args": {
      "External id": 2514,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914293.320, "dur": 6.155,
    "args": {
      "External id": 2515,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914300.964, "dur": 6.128,
    "args": {
      "External id": 2516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914308.827, "dur": 6.117,
    "args": {
      "External id": 2517,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914317.284, "dur": 7.296,
    "args": {
      "External id": 2518,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914317.947, "dur": 0.090,
    "args": {
      "External id": 2519,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740914326.285, "dur": 7.179,
    "args": {
      "External id": 2520,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914334.634, "dur": 6.354,
    "args": {
      "External id": 2521,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740914367.604, "dur": 7.796,
    "args": {
      "External id": 2522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914368.528, "dur": 6.606,
    "args": {
      "External id": 2523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914379.134, "dur": 9.132,
    "args": {
      "External id": 2524,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914380.120, "dur": 1.402,
    "args": {
      "External id": 2525,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914382.144, "dur": 5.974,
    "args": {
      "External id": 2526,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914411.563, "dur": 0.173,
    "args": {
      "External id": 2527,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914417.519, "dur": 83.700,
    "args": {
      "External id": 2528,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914425.109, "dur": 2.150,
    "args": {
      "External id": 2529,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914426.285, "dur": 0.452,
    "args": {
      "External id": 2530,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914428.609, "dur": 1.421,
    "args": {
      "External id": 2531,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914429.443, "dur": 0.269,
    "args": {
      "External id": 2532,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914432.210, "dur": 6.822,
    "args": {
      "External id": 2533,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914441.409, "dur": 7.519,
    "args": {
      "External id": 2534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914451.149, "dur": 5.796,
    "args": {
      "External id": 2535,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914458.384, "dur": 6.165,
    "args": {
      "External id": 2536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914466.259, "dur": 7.079,
    "args": {
      "External id": 2537,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914475.679, "dur": 6.932,
    "args": {
      "External id": 2538,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914476.339, "dur": 0.091,
    "args": {
      "External id": 2539,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740914484.196, "dur": 6.187,
    "args": {
      "External id": 2540,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914491.479, "dur": 6.164,
    "args": {
      "External id": 2541,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740914522.146, "dur": 6.882,
    "args": {
      "External id": 2542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914523.058, "dur": 5.690,
    "args": {
      "External id": 2543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914532.785, "dur": 9.685,
    "args": {
      "External id": 2544,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914533.734, "dur": 1.428,
    "args": {
      "External id": 2545,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914535.723, "dur": 6.579,
    "args": {
      "External id": 2546,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914564.052, "dur": 0.168,
    "args": {
      "External id": 2547,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914569.549, "dur": 90.791,
    "args": {
      "External id": 2548,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914577.027, "dur": 2.194,
    "args": {
      "External id": 2549,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914578.282, "dur": 0.437,
    "args": {
      "External id": 2550,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914580.511, "dur": 1.402,
    "args": {
      "External id": 2551,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914581.317, "dur": 0.271,
    "args": {
      "External id": 2552,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914584.121, "dur": 6.209,
    "args": {
      "External id": 2553,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914592.608, "dur": 7.004,
    "args": {
      "External id": 2554,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914601.726, "dur": 7.150,
    "args": {
      "External id": 2555,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914610.402, "dur": 6.139,
    "args": {
      "External id": 2556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914618.187, "dur": 6.420,
    "args": {
      "External id": 2557,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914626.898, "dur": 6.616,
    "args": {
      "External id": 2558,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914627.590, "dur": 0.102,
    "args": {
      "External id": 2559,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740914635.042, "dur": 6.674,
    "args": {
      "External id": 2560,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914651.133, "dur": 5.987,
    "args": {
      "External id": 2561,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740914681.037, "dur": 13.951,
    "args": {
      "External id": 2562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914681.930, "dur": 12.532,
    "args": {
      "External id": 2563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914700.713, "dur": 13.680,
    "args": {
      "External id": 2564,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914702.317, "dur": 2.325,
    "args": {
      "External id": 2565,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914705.183, "dur": 9.021,
    "args": {
      "External id": 2566,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914741.300, "dur": 0.166,
    "args": {
      "External id": 2567,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914748.307, "dur": 99.967,
    "args": {
      "External id": 2568,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914758.006, "dur": 2.923,
    "args": {
      "External id": 2569,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914759.739, "dur": 0.579,
    "args": {
      "External id": 2570,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914762.214, "dur": 1.544,
    "args": {
      "External id": 2571,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914763.143, "dur": 0.295,
    "args": {
      "External id": 2572,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914766.075, "dur": 8.266,
    "args": {
      "External id": 2573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914777.168, "dur": 7.875,
    "args": {
      "External id": 2574,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914787.447, "dur": 6.489,
    "args": {
      "External id": 2575,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914795.458, "dur": 5.722,
    "args": {
      "External id": 2576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914802.949, "dur": 7.398,
    "args": {
      "External id": 2577,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914812.806, "dur": 9.849,
    "args": {
      "External id": 2578,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914814.177, "dur": 0.113,
    "args": {
      "External id": 2579,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740914824.377, "dur": 5.991,
    "args": {
      "External id": 2580,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914831.395, "dur": 11.431,
    "args": {
      "External id": 2581,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740914874.067, "dur": 8.269,
    "args": {
      "External id": 2582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740914875.389, "dur": 6.689,
    "args": {
      "External id": 2583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740914886.422, "dur": 10.610,
    "args": {
      "External id": 2584,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914887.676, "dur": 1.865,
    "args": {
      "External id": 2585,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740914890.068, "dur": 6.804,
    "args": {
      "External id": 2586,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914920.525, "dur": 0.274,
    "args": {
      "External id": 2587,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914927.166, "dur": 90.364,
    "args": {
      "External id": 2588,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914935.803, "dur": 2.589,
    "args": {
      "External id": 2589,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914937.179, "dur": 0.548,
    "args": {
      "External id": 2590,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740914939.745, "dur": 1.442,
    "args": {
      "External id": 2591,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740914940.578, "dur": 0.268,
    "args": {
      "External id": 2592,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740914943.525, "dur": 8.216,
    "args": {
      "External id": 2593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914954.558, "dur": 7.982,
    "args": {
      "External id": 2594,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740914964.951, "dur": 6.266,
    "args": {
      "External id": 2595,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740914972.757, "dur": 6.856,
    "args": {
      "External id": 2596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740914981.404, "dur": 7.370,
    "args": {
      "External id": 2597,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740914991.074, "dur": 7.372,
    "args": {
      "External id": 2598,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740914991.843, "dur": 0.084,
    "args": {
      "External id": 2599,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915000.087, "dur": 6.772,
    "args": {
      "External id": 2600,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915007.949, "dur": 6.000,
    "args": {
      "External id": 2601,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915040.741, "dur": 6.597,
    "args": {
      "External id": 2602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915041.694, "dur": 5.373,
    "args": {
      "External id": 2603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915051.085, "dur": 9.576,
    "args": {
      "External id": 2604,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915052.047, "dur": 1.424,
    "args": {
      "External id": 2605,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915053.945, "dur": 6.583,
    "args": {
      "External id": 2606,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915082.763, "dur": 0.164,
    "args": {
      "External id": 2607,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915088.695, "dur": 84.529,
    "args": {
      "External id": 2608,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915096.389, "dur": 2.144,
    "args": {
      "External id": 2609,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915097.608, "dur": 0.412,
    "args": {
      "External id": 2610,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915099.906, "dur": 1.426,
    "args": {
      "External id": 2611,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915100.743, "dur": 0.281,
    "args": {
      "External id": 2612,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915103.471, "dur": 6.173,
    "args": {
      "External id": 2613,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915111.782, "dur": 6.908,
    "args": {
      "External id": 2614,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915121.156, "dur": 6.472,
    "args": {
      "External id": 2615,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915129.064, "dur": 6.005,
    "args": {
      "External id": 2616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915136.818, "dur": 7.333,
    "args": {
      "External id": 2617,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915146.416, "dur": 6.969,
    "args": {
      "External id": 2618,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915147.147, "dur": 0.102,
    "args": {
      "External id": 2619,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915154.925, "dur": 7.480,
    "args": {
      "External id": 2620,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915163.469, "dur": 6.260,
    "args": {
      "External id": 2621,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915194.578, "dur": 7.069,
    "args": {
      "External id": 2622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915195.485, "dur": 5.884,
    "args": {
      "External id": 2623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915205.397, "dur": 9.004,
    "args": {
      "External id": 2624,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915206.284, "dur": 1.453,
    "args": {
      "External id": 2625,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915208.277, "dur": 5.975,
    "args": {
      "External id": 2626,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915236.599, "dur": 0.169,
    "args": {
      "External id": 2627,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915242.557, "dur": 108.702,
    "args": {
      "External id": 2628,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915250.219, "dur": 2.204,
    "args": {
      "External id": 2629,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915251.423, "dur": 0.462,
    "args": {
      "External id": 2630,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915253.730, "dur": 1.403,
    "args": {
      "External id": 2631,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915254.531, "dur": 0.274,
    "args": {
      "External id": 2632,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915257.466, "dur": 6.316,
    "args": {
      "External id": 2633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915265.924, "dur": 9.248,
    "args": {
      "External id": 2634,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915277.468, "dur": 17.495,
    "args": {
      "External id": 2635,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915297.891, "dur": 9.775,
    "args": {
      "External id": 2636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915309.989, "dur": 8.994,
    "args": {
      "External id": 2637,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915321.734, "dur": 8.235,
    "args": {
      "External id": 2638,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915322.553, "dur": 0.215,
    "args": {
      "External id": 2639,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915331.720, "dur": 7.197,
    "args": {
      "External id": 2640,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915340.527, "dur": 6.636,
    "args": {
      "External id": 2641,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915376.008, "dur": 6.947,
    "args": {
      "External id": 2642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915376.939, "dur": 5.773,
    "args": {
      "External id": 2643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915386.834, "dur": 9.663,
    "args": {
      "External id": 2644,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915388.127, "dur": 1.559,
    "args": {
      "External id": 2645,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915390.223, "dur": 6.145,
    "args": {
      "External id": 2646,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915421.137, "dur": 0.169,
    "args": {
      "External id": 2647,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915427.801, "dur": 85.580,
    "args": {
      "External id": 2648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915436.260, "dur": 2.470,
    "args": {
      "External id": 2649,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915437.595, "dur": 0.509,
    "args": {
      "External id": 2650,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915440.073, "dur": 1.459,
    "args": {
      "External id": 2651,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915440.888, "dur": 0.331,
    "args": {
      "External id": 2652,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915443.828, "dur": 8.283,
    "args": {
      "External id": 2653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915454.886, "dur": 6.788,
    "args": {
      "External id": 2654,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915464.069, "dur": 6.183,
    "args": {
      "External id": 2655,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915471.743, "dur": 6.020,
    "args": {
      "External id": 2656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915479.512, "dur": 6.237,
    "args": {
      "External id": 2657,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915488.008, "dur": 6.537,
    "args": {
      "External id": 2658,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915488.669, "dur": 0.098,
    "args": {
      "External id": 2659,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915496.108, "dur": 6.395,
    "args": {
      "External id": 2660,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915503.556, "dur": 6.229,
    "args": {
      "External id": 2661,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915536.147, "dur": 7.150,
    "args": {
      "External id": 2662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915537.072, "dur": 5.970,
    "args": {
      "External id": 2663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915547.041, "dur": 9.111,
    "args": {
      "External id": 2664,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915548.005, "dur": 1.472,
    "args": {
      "External id": 2665,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915549.981, "dur": 6.014,
    "args": {
      "External id": 2666,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915579.176, "dur": 0.140,
    "args": {
      "External id": 2667,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915585.145, "dur": 81.721,
    "args": {
      "External id": 2668,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915592.784, "dur": 2.208,
    "args": {
      "External id": 2669,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915593.990, "dur": 0.442,
    "args": {
      "External id": 2670,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915596.349, "dur": 1.376,
    "args": {
      "External id": 2671,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915597.141, "dur": 0.263,
    "args": {
      "External id": 2672,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915599.976, "dur": 7.207,
    "args": {
      "External id": 2673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915609.382, "dur": 6.901,
    "args": {
      "External id": 2674,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915618.482, "dur": 5.965,
    "args": {
      "External id": 2675,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915625.825, "dur": 5.591,
    "args": {
      "External id": 2676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915633.153, "dur": 6.517,
    "args": {
      "External id": 2677,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915641.992, "dur": 7.073,
    "args": {
      "External id": 2678,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915642.666, "dur": 0.099,
    "args": {
      "External id": 2679,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915650.611, "dur": 5.990,
    "args": {
      "External id": 2680,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915657.709, "dur": 5.740,
    "args": {
      "External id": 2681,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915689.219, "dur": 5.924,
    "args": {
      "External id": 2682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915690.089, "dur": 4.828,
    "args": {
      "External id": 2683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915698.771, "dur": 8.775,
    "args": {
      "External id": 2684,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915699.688, "dur": 1.430,
    "args": {
      "External id": 2685,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915701.591, "dur": 5.781,
    "args": {
      "External id": 2686,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915730.447, "dur": 0.124,
    "args": {
      "External id": 2687,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915736.620, "dur": 82.475,
    "args": {
      "External id": 2688,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915744.088, "dur": 2.170,
    "args": {
      "External id": 2689,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915745.309, "dur": 0.439,
    "args": {
      "External id": 2690,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915747.533, "dur": 1.403,
    "args": {
      "External id": 2691,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915748.338, "dur": 0.281,
    "args": {
      "External id": 2692,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915751.090, "dur": 6.849,
    "args": {
      "External id": 2693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915760.079, "dur": 7.079,
    "args": {
      "External id": 2694,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915769.327, "dur": 5.723,
    "args": {
      "External id": 2695,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915776.452, "dur": 6.620,
    "args": {
      "External id": 2696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915784.812, "dur": 7.018,
    "args": {
      "External id": 2697,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915794.122, "dur": 7.041,
    "args": {
      "External id": 2698,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915794.720, "dur": 0.092,
    "args": {
      "External id": 2699,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915802.702, "dur": 6.356,
    "args": {
      "External id": 2700,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915810.120, "dur": 5.709,
    "args": {
      "External id": 2701,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740915840.694, "dur": 6.951,
    "args": {
      "External id": 2702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740915841.605, "dur": 5.740,
    "args": {
      "External id": 2703,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740915851.275, "dur": 14.277,
    "args": {
      "External id": 2704,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915852.210, "dur": 1.324,
    "args": {
      "External id": 2705,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740915854.050, "dur": 11.012,
    "args": {
      "External id": 2706,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915894.170, "dur": 0.342,
    "args": {
      "External id": 2707,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915902.710, "dur": 104.754,
    "args": {
      "External id": 2708,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915911.276, "dur": 3.328,
    "args": {
      "External id": 2709,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915913.288, "dur": 0.646,
    "args": {
      "External id": 2710,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740915915.969, "dur": 1.638,
    "args": {
      "External id": 2711,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740915916.966, "dur": 0.303,
    "args": {
      "External id": 2712,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740915919.878, "dur": 10.634,
    "args": {
      "External id": 2713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915933.366, "dur": 8.640,
    "args": {
      "External id": 2714,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915944.568, "dur": 7.337,
    "args": {
      "External id": 2715,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740915953.392, "dur": 7.009,
    "args": {
      "External id": 2716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740915962.169, "dur": 7.308,
    "args": {
      "External id": 2717,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740915971.971, "dur": 7.624,
    "args": {
      "External id": 2718,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740915972.699, "dur": 0.132,
    "args": {
      "External id": 2719,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740915981.291, "dur": 7.164,
    "args": {
      "External id": 2720,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740915989.610, "dur": 12.739,
    "args": {
      "External id": 2721,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916034.809, "dur": 7.732,
    "args": {
      "External id": 2722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916036.084, "dur": 6.157,
    "args": {
      "External id": 2723,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916046.602, "dur": 11.016,
    "args": {
      "External id": 2724,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916047.827, "dur": 1.833,
    "args": {
      "External id": 2725,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916050.188, "dur": 7.248,
    "args": {
      "External id": 2726,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916081.799, "dur": 0.153,
    "args": {
      "External id": 2727,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916088.452, "dur": 88.622,
    "args": {
      "External id": 2728,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916096.815, "dur": 2.523,
    "args": {
      "External id": 2729,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916098.208, "dur": 0.510,
    "args": {
      "External id": 2730,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916100.792, "dur": 1.448,
    "args": {
      "External id": 2731,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916101.650, "dur": 0.285,
    "args": {
      "External id": 2732,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916104.620, "dur": 6.954,
    "args": {
      "External id": 2733,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916114.400, "dur": 7.900,
    "args": {
      "External id": 2734,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916124.760, "dur": 6.683,
    "args": {
      "External id": 2735,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916132.938, "dur": 6.808,
    "args": {
      "External id": 2736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916141.494, "dur": 6.403,
    "args": {
      "External id": 2737,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916150.291, "dur": 7.144,
    "args": {
      "External id": 2738,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916150.992, "dur": 0.087,
    "args": {
      "External id": 2739,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916159.172, "dur": 7.127,
    "args": {
      "External id": 2740,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916167.484, "dur": 5.987,
    "args": {
      "External id": 2741,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916199.869, "dur": 6.859,
    "args": {
      "External id": 2742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916200.742, "dur": 5.718,
    "args": {
      "External id": 2743,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916210.368, "dur": 10.324,
    "args": {
      "External id": 2744,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916211.280, "dur": 1.540,
    "args": {
      "External id": 2745,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916213.365, "dur": 7.171,
    "args": {
      "External id": 2746,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916243.461, "dur": 0.140,
    "args": {
      "External id": 2747,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916249.587, "dur": 82.730,
    "args": {
      "External id": 2748,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916257.308, "dur": 2.178,
    "args": {
      "External id": 2749,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916258.526, "dur": 0.412,
    "args": {
      "External id": 2750,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916260.862, "dur": 1.458,
    "args": {
      "External id": 2751,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916261.718, "dur": 0.267,
    "args": {
      "External id": 2752,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916264.512, "dur": 7.435,
    "args": {
      "External id": 2753,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916274.424, "dur": 6.948,
    "args": {
      "External id": 2754,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916283.591, "dur": 5.916,
    "args": {
      "External id": 2755,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916290.964, "dur": 6.025,
    "args": {
      "External id": 2756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916298.746, "dur": 6.725,
    "args": {
      "External id": 2757,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916307.803, "dur": 6.637,
    "args": {
      "External id": 2758,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916308.430, "dur": 0.095,
    "args": {
      "External id": 2759,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916316.056, "dur": 5.837,
    "args": {
      "External id": 2760,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916322.962, "dur": 5.948,
    "args": {
      "External id": 2761,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916354.349, "dur": 6.713,
    "args": {
      "External id": 2762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916355.171, "dur": 5.648,
    "args": {
      "External id": 2763,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916364.742, "dur": 8.923,
    "args": {
      "External id": 2764,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916365.720, "dur": 1.481,
    "args": {
      "External id": 2765,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916367.718, "dur": 5.770,
    "args": {
      "External id": 2766,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916396.984, "dur": 0.118,
    "args": {
      "External id": 2767,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916402.615, "dur": 103.117,
    "args": {
      "External id": 2768,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916410.568, "dur": 2.178,
    "args": {
      "External id": 2769,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916411.802, "dur": 0.399,
    "args": {
      "External id": 2770,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916414.042, "dur": 1.429,
    "args": {
      "External id": 2771,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916414.853, "dur": 0.286,
    "args": {
      "External id": 2772,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916417.544, "dur": 6.652,
    "args": {
      "External id": 2773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916426.429, "dur": 6.708,
    "args": {
      "External id": 2774,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916435.237, "dur": 6.171,
    "args": {
      "External id": 2775,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916442.819, "dur": 5.474,
    "args": {
      "External id": 2776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916450.034, "dur": 9.035,
    "args": {
      "External id": 2777,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916461.283, "dur": 17.823,
    "args": {
      "External id": 2778,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916461.903, "dur": 0.174,
    "args": {
      "External id": 2779,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916482.911, "dur": 9.985,
    "args": {
      "External id": 2780,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916494.224, "dur": 7.481,
    "args": {
      "External id": 2781,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916532.021, "dur": 7.039,
    "args": {
      "External id": 2782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916532.996, "dur": 5.829,
    "args": {
      "External id": 2783,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916543.076, "dur": 10.022,
    "args": {
      "External id": 2784,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916544.269, "dur": 1.676,
    "args": {
      "External id": 2785,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916546.478, "dur": 6.468,
    "args": {
      "External id": 2786,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916576.989, "dur": 0.232,
    "args": {
      "External id": 2787,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916583.725, "dur": 86.922,
    "args": {
      "External id": 2788,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916592.937, "dur": 2.596,
    "args": {
      "External id": 2789,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916594.416, "dur": 0.479,
    "args": {
      "External id": 2790,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916596.870, "dur": 1.354,
    "args": {
      "External id": 2791,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916597.642, "dur": 0.277,
    "args": {
      "External id": 2792,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916600.597, "dur": 7.016,
    "args": {
      "External id": 2793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916610.326, "dur": 6.676,
    "args": {
      "External id": 2794,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916619.328, "dur": 6.977,
    "args": {
      "External id": 2795,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916627.893, "dur": 5.772,
    "args": {
      "External id": 2796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916635.338, "dur": 6.991,
    "args": {
      "External id": 2797,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916644.805, "dur": 6.753,
    "args": {
      "External id": 2798,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916645.539, "dur": 0.089,
    "args": {
      "External id": 2799,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916653.122, "dur": 6.674,
    "args": {
      "External id": 2800,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916660.946, "dur": 6.056,
    "args": {
      "External id": 2801,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916692.826, "dur": 8.278,
    "args": {
      "External id": 2802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916693.721, "dur": 7.078,
    "args": {
      "External id": 2803,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916705.021, "dur": 9.470,
    "args": {
      "External id": 2804,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916705.973, "dur": 1.534,
    "args": {
      "External id": 2805,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916708.010, "dur": 6.285,
    "args": {
      "External id": 2806,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916737.228, "dur": 0.130,
    "args": {
      "External id": 2807,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916743.488, "dur": 83.196,
    "args": {
      "External id": 2808,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916751.335, "dur": 2.178,
    "args": {
      "External id": 2809,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916752.526, "dur": 0.430,
    "args": {
      "External id": 2810,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916754.819, "dur": 1.426,
    "args": {
      "External id": 2811,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916755.652, "dur": 0.275,
    "args": {
      "External id": 2812,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916758.534, "dur": 6.250,
    "args": {
      "External id": 2813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916767.053, "dur": 7.900,
    "args": {
      "External id": 2814,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916777.140, "dur": 6.348,
    "args": {
      "External id": 2815,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916785.007, "dur": 6.138,
    "args": {
      "External id": 2816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916792.799, "dur": 6.580,
    "args": {
      "External id": 2817,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916801.768, "dur": 6.617,
    "args": {
      "External id": 2818,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916802.363, "dur": 0.096,
    "args": {
      "External id": 2819,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916809.936, "dur": 5.945,
    "args": {
      "External id": 2820,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916816.915, "dur": 6.262,
    "args": {
      "External id": 2821,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740916848.194, "dur": 6.728,
    "args": {
      "External id": 2822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740916849.014, "dur": 5.643,
    "args": {
      "External id": 2823,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740916858.692, "dur": 10.433,
    "args": {
      "External id": 2824,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916859.614, "dur": 1.391,
    "args": {
      "External id": 2825,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740916861.538, "dur": 7.436,
    "args": {
      "External id": 2826,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916890.653, "dur": 0.147,
    "args": {
      "External id": 2827,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916896.833, "dur": 85.354,
    "args": {
      "External id": 2828,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916904.590, "dur": 2.379,
    "args": {
      "External id": 2829,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916905.903, "dur": 0.508,
    "args": {
      "External id": 2830,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740916908.299, "dur": 1.523,
    "args": {
      "External id": 2831,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740916909.159, "dur": 0.330,
    "args": {
      "External id": 2832,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740916912.025, "dur": 8.162,
    "args": {
      "External id": 2833,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916922.568, "dur": 7.289,
    "args": {
      "External id": 2834,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916931.915, "dur": 6.428,
    "args": {
      "External id": 2835,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740916939.823, "dur": 6.303,
    "args": {
      "External id": 2836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740916947.770, "dur": 6.355,
    "args": {
      "External id": 2837,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740916956.505, "dur": 6.846,
    "args": {
      "External id": 2838,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740916957.110, "dur": 0.087,
    "args": {
      "External id": 2839,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740916964.943, "dur": 6.868,
    "args": {
      "External id": 2840,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740916972.826, "dur": 6.094,
    "args": {
      "External id": 2841,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740917003.489, "dur": 7.053,
    "args": {
      "External id": 2842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740917004.459, "dur": 5.860,
    "args": {
      "External id": 2843,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740917014.160, "dur": 8.967,
    "args": {
      "External id": 2844,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917015.117, "dur": 1.425,
    "args": {
      "External id": 2845,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740917017.039, "dur": 5.939,
    "args": {
      "External id": 2846,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917045.061, "dur": 0.116,
    "args": {
      "External id": 2847,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917056.280, "dur": 120.462,
    "args": {
      "External id": 2848,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917067.114, "dur": 5.751,
    "args": {
      "External id": 2849,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917071.119, "dur": 0.876,
    "args": {
      "External id": 2850,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917074.500, "dur": 2.410,
    "args": {
      "External id": 2851,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917075.676, "dur": 0.836,
    "args": {
      "External id": 2852,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740917079.547, "dur": 9.548,
    "args": {
      "External id": 2853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917092.213, "dur": 10.524,
    "args": {
      "External id": 2854,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917105.549, "dur": 7.872,
    "args": {
      "External id": 2855,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740917114.876, "dur": 7.217,
    "args": {
      "External id": 2856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740917123.815, "dur": 21.605,
    "args": {
      "External id": 2857,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740917148.112, "dur": 8.224,
    "args": {
      "External id": 2858,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917148.869, "dur": 0.119,
    "args": {
      "External id": 2859,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740917157.969, "dur": 7.431,
    "args": {
      "External id": 2860,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917166.596, "dur": 6.236,
    "args": {
      "External id": 2861,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740917202.805, "dur": 9.222,
    "args": {
      "External id": 2862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740917203.955, "dur": 7.800,
    "args": {
      "External id": 2863,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740917216.026, "dur": 10.429,
    "args": {
      "External id": 2864,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917217.591, "dur": 1.582,
    "args": {
      "External id": 2865,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740917219.729, "dur": 6.581,
    "args": {
      "External id": 2866,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917251.249, "dur": 0.148,
    "args": {
      "External id": 2867,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917257.963, "dur": 87.733,
    "args": {
      "External id": 2868,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917266.003, "dur": 2.326,
    "args": {
      "External id": 2869,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917267.326, "dur": 0.445,
    "args": {
      "External id": 2870,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917269.666, "dur": 1.409,
    "args": {
      "External id": 2871,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917270.447, "dur": 0.279,
    "args": {
      "External id": 2872,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740917273.318, "dur": 6.443,
    "args": {
      "External id": 2873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917282.245, "dur": 8.147,
    "args": {
      "External id": 2874,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917292.871, "dur": 6.513,
    "args": {
      "External id": 2875,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740917300.796, "dur": 6.109,
    "args": {
      "External id": 2876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740917308.523, "dur": 6.854,
    "args": {
      "External id": 2877,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740917317.792, "dur": 7.002,
    "args": {
      "External id": 2878,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917318.486, "dur": 0.104,
    "args": {
      "External id": 2879,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740917326.456, "dur": 8.006,
    "args": {
      "External id": 2880,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917335.641, "dur": 6.564,
    "args": {
      "External id": 2881,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 49625, "tid": 49625,
    "ts": 7253740917368.645, "dur": 7.272,
    "args": {
      "External id": 2882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Strides": [[1]], "finished": true, "Input Dims": [[67108864]], "Ev Idx": 2881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 49625, "tid": 49625,
    "ts": 7253740917369.597, "dur": 6.086,
    "args": {
      "External id": 2883,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[67108864], []], "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 49625, "tid": 49625,
    "ts": 7253740917379.641, "dur": 9.314,
    "args": {
      "External id": 2884,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], []], "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917380.534, "dur": 1.470,
    "args": {
      "External id": 2885,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "6", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "finished": true, "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 49625, "tid": 49625,
    "ts": 7253740917382.723, "dur": 6.086,
    "args": {
      "External id": 2886,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 2885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917411.040, "dur": 0.144,
    "args": {
      "External id": 2887,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "finished": true, "Input Dims": [[4096], [], [], [], []], "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "quark::scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917416.710, "dur": 83.373,
    "args": {
      "External id": 2888,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "0", "1", "0.", "255.", "8", "", ""], "Input type": ["", "float", "float", "int", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [8192, 1], [1], [1], [], [], [], [], [], [], []], "finished": true, "Input Dims": [[], [4096, 8192], [4096], [4096], [], [], [], [], [], [], []], "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917424.142, "dur": 2.177,
    "args": {
      "External id": 2889,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917425.380, "dur": 0.399,
    "args": {
      "External id": 2890,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 49625, "tid": 49625,
    "ts": 7253740917427.660, "dur": 1.467,
    "args": {
      "External id": 2891,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["int", "Scalar"], "Input Strides": [[1], []], "finished": true, "Input Dims": [[4096], []], "Ev Idx": 2890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 49625, "tid": 49625,
    "ts": 7253740917428.537, "dur": 0.282,
    "args": {
      "External id": 2892,"Record function id": 0, "Concrete Inputs": ["", "[4096, 1]", "[1, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "finished": true, "Input Dims": [[4096], [], [], []], "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 49625, "tid": 49625,
    "ts": 7253740917431.315, "dur": 7.880,
    "args": {
      "External id": 2893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "finished": true, "Input Dims": [[4096, 1]], "Ev Idx": 2892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917441.604, "dur": 6.729,
    "args": {
      "External id": 2894,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[1, 1], []], "finished": true, "Input Dims": [[4096, 1], []], "Ev Idx": 2893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917450.554, "dur": 6.240,
    "args": {
      "External id": 2895,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 49625, "tid": 49625,
    "ts": 7253740917458.218, "dur": 6.197,
    "args": {
      "External id": 2896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[8192, 1]], "finished": true, "Input Dims": [[4096, 8192]], "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 49625, "tid": 49625,
    "ts": 7253740917466.065, "dur": 6.267,
    "args": {
      "External id": 2897,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 49625, "tid": 49625,
    "ts": 7253740917474.718, "dur": 7.312,
    "args": {
      "External id": 2898,"Record function id": 0, "Concrete Inputs": ["", "0", "255"], "Input type": ["float", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], []], "finished": true, "Input Dims": [[4096, 8192], [], []], "Ev Idx": 2897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 49625, "tid": 49625,
    "ts": 7253740917475.333, "dur": 0.089,
    "args": {
      "External id": 2899,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "finished": true, "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 49625, "tid": 49625,
    "ts": 7253740917483.734, "dur": 6.035,
    "args": {
      "External id": 2900,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "int", "Scalar"], "Input Strides": [[8192, 1], [1, 1], []], "finished": true, "Input Dims": [[4096, 8192], [4096, 1], []], "Ev Idx": 2899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 49625, "tid": 49625,
    "ts": 7253740917490.781, "dur": 5.854,
    "args": {
      "External id": 2901,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[8192, 1], [1, 1]], "finished": true, "Input Dims": [[4096, 8192], [4096, 1]], "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1032): _bootstrap", "pid": 0, "tid": 0,
    "ts": 7253740836672.062, "dur": 124912.389,
    "args": {
      "Python parent id": null, "Python id": 1, "finished": true, "Ev Idx": 2901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1075): _bootstrap_inner", "pid": 0, "tid": 0,
    "ts": 7253740836673.775, "dur": 124910.676,
    "args": {
      "Python parent id": 1, "Python id": 2, "finished": true, "Ev Idx": 2902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "tqdm/_monitor.py(60): run", "pid": 0, "tid": 0,
    "ts": 7253740836675.867, "dur": 124908.584,
    "args": {
      "Python parent id": 2, "Python id": 3, "finished": true, "Ev Idx": 2903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(655): wait", "pid": 0, "tid": 0,
    "ts": 7253740836676.796, "dur": 124907.655,
    "args": {
      "Python parent id": 3, "Python id": 4, "finished": true, "Ev Idx": 2904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(359): wait", "pid": 0, "tid": 0,
    "ts": 7253740836678.814, "dur": 124905.637,
    "args": {
      "Python parent id": 4, "Python id": 5, "finished": true, "Ev Idx": 2905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1032): _bootstrap", "pid": 0, "tid": 0,
    "ts": 7253740836713.375, "dur": 124871.076,
    "args": {
      "Python parent id": 5, "Python id": 6, "finished": true, "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1075): _bootstrap_inner", "pid": 0, "tid": 0,
    "ts": 7253740836713.846, "dur": 124870.605,
    "args": {
      "Python parent id": 6, "Python id": 7, "finished": true, "Ev Idx": 2907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1012): run", "pid": 0, "tid": 0,
    "ts": 7253740836714.722, "dur": 124869.729,
    "args": {
      "Python parent id": 7, "Python id": 8, "finished": true, "Ev Idx": 2908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_inductor/compile_worker/subproc_pool.py(191): _read_thread", "pid": 0, "tid": 0,
    "ts": 7253740836715.534, "dur": 124868.917,
    "args": {
      "Python parent id": 8, "Python id": 9, "finished": true, "Ev Idx": 2909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_inductor/compile_worker/subproc_pool.py(55): _recv_msg", "pid": 0, "tid": 0,
    "ts": 7253740836716.496, "dur": 124867.955,
    "args": {
      "Python parent id": 9, "Python id": 10, "finished": true, "Ev Idx": 2910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(206): <module>", "pid": 49625, "tid": 49625,
    "ts": 7253740836725.787, "dur": 124858.664,
    "args": {
      "Python parent id": null, "Python id": 11, "finished": true, "Ev Idx": 2911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(113): do_bench", "pid": 49625, "tid": 49625,
    "ts": 7253740836727.453, "dur": 124856.998,
    "args": {
      "Python parent id": 11, "Python id": 12, "finished": true, "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(800): __enter__", "pid": 49625, "tid": 49625,
    "ts": 7253740836728.104, "dur": 93.979,
    "args": {
      "Python parent id": 12, "Python id": 13, "finished": true, "Ev Idx": 2913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(810): start", "pid": 49625, "tid": 49625,
    "ts": 7253740836728.677, "dur": 93.217,
    "args": {
      "Python parent id": 13, "Python id": 14, "finished": true, "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(866): _transit_action", "pid": 49625, "tid": 49625,
    "ts": 7253740836729.500, "dur": 91.998,
    "args": {
      "Python parent id": 14, "Python id": 15, "finished": true, "Ev Idx": 2915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(203): start_trace", "pid": 49625, "tid": 49625,
    "ts": 7253740836730.412, "dur": 90.661,
    "args": {
      "Python parent id": 15, "Python id": 16, "finished": true, "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/profiler.py(370): _start_trace", "pid": 49625, "tid": 49625,
    "ts": 7253740836731.205, "dur": 26.243,
    "args": {
      "Python parent id": 16, "Python id": 17, "finished": true, "Ev Idx": 2917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function perf_counter_ns>", "pid": 49625, "tid": 49625,
    "ts": 7253740836754.041, "dur": 1.214,
    "args": {
      "Python parent id": 17, "Python id": 18, "finished": true, "Ev Idx": 2918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(340): add_metadata_json", "pid": 49625, "tid": 49625,
    "ts": 7253740836762.065, "dur": 8.684,
    "args": {
      "Python parent id": 16, "Python id": 19, "finished": true, "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _add_metadata_json of PyCapsule object at 0x7f76b136fa20>", "pid": 49625, "tid": 49625,
    "ts": 7253740836768.083, "dur": 2.454,
    "args": {
      "Python parent id": 19, "Python id": 20, "finished": true, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(340): add_metadata_json", "pid": 49625, "tid": 49625,
    "ts": 7253740836772.211, "dur": 1.384,
    "args": {
      "Python parent id": 16, "Python id": 21, "finished": true, "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _add_metadata_json of PyCapsule object at 0x7f76b136fa20>", "pid": 49625, "tid": 49625,
    "ts": 7253740836773.027, "dur": 0.502,
    "args": {
      "Python parent id": 21, "Python id": 22, "finished": true, "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method kineto_available of PyCapsule object at 0x7f76b136fa80>", "pid": 49625, "tid": 49625,
    "ts": 7253740836776.828, "dur": 0.608,
    "args": {
      "Python parent id": 16, "Python id": 23, "finished": true, "Ev Idx": 2923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(355): _get_distributed_info", "pid": 49625, "tid": 49625,
    "ts": 7253740836781.186, "dur": 12.089,
    "args": {
      "Python parent id": 16, "Python id": 24, "finished": true, "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/__init__.py(14): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740836785.138, "dur": 2.739,
    "args": {
      "Python parent id": 24, "Python id": 25, "finished": true, "Ev Idx": 2925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740836786.753, "dur": 0.976,
    "args": {
      "Python parent id": 25, "Python id": 26, "finished": true, "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(1269): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740836789.283, "dur": 3.809,
    "args": {
      "Python parent id": 24, "Python id": 27, "finished": true, "Ev Idx": 2927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(721): WORLD", "pid": 49625, "tid": 49625,
    "ts": 7253740836791.438, "dur": 1.524,
    "args": {
      "Python parent id": 27, "Python id": 28, "finished": true, "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(596): default_pg", "pid": 49625, "tid": 49625,
    "ts": 7253740836792.566, "dur": 0.282,
    "args": {
      "Python parent id": 28, "Python id": 29, "finished": true, "Ev Idx": 2929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740836794.979, "dur": 0.405,
    "args": {
      "Python parent id": 16, "Python id": 30, "finished": true, "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(774): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740836802.212, "dur": 14.977,
    "args": {
      "Python parent id": 16, "Python id": 31, "finished": true, "Ev Idx": 2931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(369): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740836805.538, "dur": 11.419,
    "args": {
      "Python parent id": 31, "Python id": 32, "finished": true, "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(425): _get_alias_val", "pid": 49625, "tid": 49625,
    "ts": 7253740836809.194, "dur": 3.105,
    "args": {
      "Python parent id": 32, "Python id": 33, "finished": true, "Ev Idx": 2933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(412): _get_alias_module_and_name", "pid": 49625, "tid": 49625,
    "ts": 7253740836811.459, "dur": 0.402,
    "args": {
      "Python parent id": 33, "Python id": 34, "finished": true, "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740836815.909, "dur": 0.813,
    "args": {
      "Python parent id": 32, "Python id": 35, "finished": true, "Ev Idx": 2935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740836819.911, "dur": 0.452,
    "args": {
      "Python parent id": 16, "Python id": 36, "finished": true, "Ev Idx": 2936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740836824.191, "dur": 0.160,
    "args": {
      "Python parent id": 12, "Python id": 37, "finished": true, "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740836826.981, "dur": 1.262,
    "args": {
      "Python parent id": 12, "Python id": 38, "finished": true, "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740836829.592, "dur": 0.096,
    "args": {
      "Python parent id": 12, "Python id": 39, "finished": true, "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740836831.715, "dur": 247.992,
    "args": {
      "Python parent id": 12, "Python id": 40, "finished": true, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740837084.670, "dur": 37914.527,
    "args": {
      "Python parent id": 12, "Python id": 41, "finished": true, "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740837089.449, "dur": 2.221,
    "args": {
      "Python parent id": 41, "Python id": 42, "finished": true, "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740837091.069, "dur": 0.503,
    "args": {
      "Python parent id": 42, "Python id": 43, "finished": true, "Ev Idx": 2943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740837093.652, "dur": 1.693,
    "args": {
      "Python parent id": 41, "Python id": 44, "finished": true, "Ev Idx": 2944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740837096.022, "dur": 0.342,
    "args": {
      "Python parent id": 41, "Python id": 45, "finished": true, "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740837099.037, "dur": 5.941,
    "args": {
      "Python parent id": 41, "Python id": 46, "finished": true, "Ev Idx": 2946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740837110.006, "dur": 37888.715,
    "args": {
      "Python parent id": 41, "Python id": 47, "finished": true, "Ev Idx": 2947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740837116.509, "dur": 37880.926,
    "args": {
      "Python parent id": 47, "Python id": 48, "finished": true, "Ev Idx": 2948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740837119.812, "dur": 37877.535,
    "args": {
      "Python parent id": 48, "Python id": 49, "finished": true, "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740837152.446, "dur": 37834.918,
    "args": {
      "Python parent id": 49, "Python id": 50, "finished": true, "Ev Idx": 2950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740837155.028, "dur": 37831.905,
    "args": {
      "Python parent id": 50, "Python id": 51, "finished": true, "Ev Idx": 2951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740837160.354, "dur": 37825.816,
    "args": {
      "Python parent id": 51, "Python id": 52, "finished": true, "Ev Idx": 2952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740837163.728, "dur": 1.378,
    "args": {
      "Python parent id": 52, "Python id": 53, "finished": true, "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740837164.239, "dur": 0.777,
    "args": {
      "Python parent id": 53, "Python id": 54, "finished": true, "Ev Idx": 2954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740837166.391, "dur": 0.573,
    "args": {
      "Python parent id": 52, "Python id": 55, "finished": true, "Ev Idx": 2955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740837166.716, "dur": 0.053,
    "args": {
      "Python parent id": 55, "Python id": 56, "finished": true, "Ev Idx": 2956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740837172.436, "dur": 37812.998,
    "args": {
      "Python parent id": 52, "Python id": 57, "finished": true, "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740837177.303, "dur": 37801.106,
    "args": {
      "Python parent id": 57, "Python id": 58, "finished": true, "Ev Idx": 2958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740837223.378, "dur": 13576.666,
    "args": {
      "Python parent id": 58, "Python id": 59, "finished": true, "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740837225.356, "dur": 0.284,
    "args": {
      "Python parent id": 59, "Python id": 60, "finished": true, "Ev Idx": 2960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740837227.298, "dur": 13570.926,
    "args": {
      "Python parent id": 59, "Python id": 61, "finished": true, "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740837229.231, "dur": 7783.119,
    "args": {
      "Python parent id": 61, "Python id": 62, "finished": true, "Ev Idx": 2962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740850924.049, "dur": 67.470,
    "args": {
      "Python parent id": 58, "Python id": 63, "finished": true, "Ev Idx": 2963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740861154.961, "dur": 13684.853,
    "args": {
      "Python parent id": 58, "Python id": 64, "finished": true, "Ev Idx": 2964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875004.599, "dur": 0.412,
    "args": {
      "Python parent id": 12, "Python id": 65, "finished": true, "Ev Idx": 2965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875007.279, "dur": 1.542,
    "args": {
      "Python parent id": 12, "Python id": 66, "finished": true, "Ev Idx": 2966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875009.601, "dur": 0.064,
    "args": {
      "Python parent id": 12, "Python id": 67, "finished": true, "Ev Idx": 2967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875010.377, "dur": 34.074,
    "args": {
      "Python parent id": 12, "Python id": 68, "finished": true, "Ev Idx": 2968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875049.050, "dur": 203.654,
    "args": {
      "Python parent id": 12, "Python id": 69, "finished": true, "Ev Idx": 2969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875051.332, "dur": 2.250,
    "args": {
      "Python parent id": 69, "Python id": 70, "finished": true, "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875052.934, "dur": 0.560,
    "args": {
      "Python parent id": 70, "Python id": 71, "finished": true, "Ev Idx": 2971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875054.972, "dur": 1.741,
    "args": {
      "Python parent id": 69, "Python id": 72, "finished": true, "Ev Idx": 2972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875057.227, "dur": 0.442,
    "args": {
      "Python parent id": 69, "Python id": 73, "finished": true, "Ev Idx": 2973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875058.975, "dur": 4.638,
    "args": {
      "Python parent id": 69, "Python id": 74, "finished": true, "Ev Idx": 2974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875066.699, "dur": 185.841,
    "args": {
      "Python parent id": 69, "Python id": 75, "finished": true, "Ev Idx": 2975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875071.774, "dur": 180.394,
    "args": {
      "Python parent id": 75, "Python id": 76, "finished": true, "Ev Idx": 2976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875072.972, "dur": 179.136,
    "args": {
      "Python parent id": 76, "Python id": 77, "finished": true, "Ev Idx": 2977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875096.366, "dur": 153.239,
    "args": {
      "Python parent id": 77, "Python id": 78, "finished": true, "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875097.730, "dur": 151.739,
    "args": {
      "Python parent id": 78, "Python id": 79, "finished": true, "Ev Idx": 2979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875100.936, "dur": 148.327,
    "args": {
      "Python parent id": 79, "Python id": 80, "finished": true, "Ev Idx": 2980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875104.517, "dur": 1.346,
    "args": {
      "Python parent id": 80, "Python id": 81, "finished": true, "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875105.005, "dur": 0.781,
    "args": {
      "Python parent id": 81, "Python id": 82, "finished": true, "Ev Idx": 2982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875107.095, "dur": 0.528,
    "args": {
      "Python parent id": 80, "Python id": 83, "finished": true, "Ev Idx": 2983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875107.387, "dur": 0.052,
    "args": {
      "Python parent id": 83, "Python id": 84, "finished": true, "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875109.117, "dur": 139.900,
    "args": {
      "Python parent id": 80, "Python id": 85, "finished": true, "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875111.996, "dur": 134.633,
    "args": {
      "Python parent id": 85, "Python id": 86, "finished": true, "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875136.928, "dur": 38.823,
    "args": {
      "Python parent id": 86, "Python id": 87, "finished": true, "Ev Idx": 2987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875137.975, "dur": 0.380,
    "args": {
      "Python parent id": 87, "Python id": 88, "finished": true, "Ev Idx": 2988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875139.578, "dur": 35.871,
    "args": {
      "Python parent id": 87, "Python id": 89, "finished": true, "Ev Idx": 2989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875140.145, "dur": 17.050,
    "args": {
      "Python parent id": 89, "Python id": 90, "finished": true, "Ev Idx": 2990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875185.908, "dur": 13.425,
    "args": {
      "Python parent id": 86, "Python id": 91, "finished": true, "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875213.224, "dur": 15.538,
    "args": {
      "Python parent id": 86, "Python id": 92, "finished": true, "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875253.997, "dur": 0.201,
    "args": {
      "Python parent id": 12, "Python id": 93, "finished": true, "Ev Idx": 2993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875254.901, "dur": 0.388,
    "args": {
      "Python parent id": 12, "Python id": 94, "finished": true, "Ev Idx": 2994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875255.737, "dur": 0.063,
    "args": {
      "Python parent id": 12, "Python id": 95, "finished": true, "Ev Idx": 2995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875256.147, "dur": 13.499,
    "args": {
      "Python parent id": 12, "Python id": 96, "finished": true, "Ev Idx": 2996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875271.596, "dur": 100.655,
    "args": {
      "Python parent id": 12, "Python id": 97, "finished": true, "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875272.082, "dur": 0.561,
    "args": {
      "Python parent id": 97, "Python id": 98, "finished": true, "Ev Idx": 2998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875272.488, "dur": 0.081,
    "args": {
      "Python parent id": 98, "Python id": 99, "finished": true, "Ev Idx": 2999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875273.231, "dur": 0.617,
    "args": {
      "Python parent id": 97, "Python id": 100, "finished": true, "Ev Idx": 3000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875274.200, "dur": 0.290,
    "args": {
      "Python parent id": 97, "Python id": 101, "finished": true, "Ev Idx": 3001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875274.918, "dur": 1.736,
    "args": {
      "Python parent id": 97, "Python id": 102, "finished": true, "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875277.691, "dur": 94.416,
    "args": {
      "Python parent id": 97, "Python id": 103, "finished": true, "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875279.226, "dur": 92.601,
    "args": {
      "Python parent id": 103, "Python id": 104, "finished": true, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875279.770, "dur": 92.001,
    "args": {
      "Python parent id": 104, "Python id": 105, "finished": true, "Ev Idx": 3005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875287.388, "dur": 82.643,
    "args": {
      "Python parent id": 105, "Python id": 106, "finished": true, "Ev Idx": 3006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875287.828, "dur": 82.093,
    "args": {
      "Python parent id": 106, "Python id": 107, "finished": true, "Ev Idx": 3007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875288.809, "dur": 80.896,
    "args": {
      "Python parent id": 107, "Python id": 108, "finished": true, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875289.684, "dur": 0.454,
    "args": {
      "Python parent id": 108, "Python id": 109, "finished": true, "Ev Idx": 3009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875290.012, "dur": 0.079,
    "args": {
      "Python parent id": 109, "Python id": 110, "finished": true, "Ev Idx": 3010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875290.608, "dur": 0.368,
    "args": {
      "Python parent id": 108, "Python id": 111, "finished": true, "Ev Idx": 3011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875290.874, "dur": 0.054,
    "args": {
      "Python parent id": 111, "Python id": 112, "finished": true, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875291.481, "dur": 78.061,
    "args": {
      "Python parent id": 108, "Python id": 113, "finished": true, "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875292.438, "dur": 75.475,
    "args": {
      "Python parent id": 113, "Python id": 114, "finished": true, "Ev Idx": 3014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875300.386, "dur": 19.584,
    "args": {
      "Python parent id": 114, "Python id": 115, "finished": true, "Ev Idx": 3015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875300.556, "dur": 0.172,
    "args": {
      "Python parent id": 115, "Python id": 116, "finished": true, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875301.185, "dur": 18.629,
    "args": {
      "Python parent id": 115, "Python id": 117, "finished": true, "Ev Idx": 3017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875301.489, "dur": 8.182,
    "args": {
      "Python parent id": 117, "Python id": 118, "finished": true, "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875327.770, "dur": 7.771,
    "args": {
      "Python parent id": 114, "Python id": 119, "finished": true, "Ev Idx": 3019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875344.746, "dur": 8.644,
    "args": {
      "Python parent id": 114, "Python id": 120, "finished": true, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875373.355, "dur": 0.180,
    "args": {
      "Python parent id": 12, "Python id": 121, "finished": true, "Ev Idx": 3021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875374.033, "dur": 0.193,
    "args": {
      "Python parent id": 12, "Python id": 122, "finished": true, "Ev Idx": 3022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875374.534, "dur": 0.064,
    "args": {
      "Python parent id": 12, "Python id": 123, "finished": true, "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875374.890, "dur": 9.614,
    "args": {
      "Python parent id": 12, "Python id": 124, "finished": true, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875386.082, "dur": 87.204,
    "args": {
      "Python parent id": 12, "Python id": 125, "finished": true, "Ev Idx": 3025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875386.424, "dur": 0.536,
    "args": {
      "Python parent id": 125, "Python id": 126, "finished": true, "Ev Idx": 3026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875386.845, "dur": 0.051,
    "args": {
      "Python parent id": 126, "Python id": 127, "finished": true, "Ev Idx": 3027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875387.390, "dur": 0.523,
    "args": {
      "Python parent id": 125, "Python id": 128, "finished": true, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875388.193, "dur": 0.232,
    "args": {
      "Python parent id": 125, "Python id": 129, "finished": true, "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875388.748, "dur": 1.491,
    "args": {
      "Python parent id": 125, "Python id": 130, "finished": true, "Ev Idx": 3030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875390.982, "dur": 82.171,
    "args": {
      "Python parent id": 125, "Python id": 131, "finished": true, "Ev Idx": 3031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875392.120, "dur": 80.824,
    "args": {
      "Python parent id": 131, "Python id": 132, "finished": true, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875392.515, "dur": 80.395,
    "args": {
      "Python parent id": 132, "Python id": 133, "finished": true, "Ev Idx": 3033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875397.747, "dur": 73.693,
    "args": {
      "Python parent id": 133, "Python id": 134, "finished": true, "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875398.068, "dur": 73.232,
    "args": {
      "Python parent id": 134, "Python id": 135, "finished": true, "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875398.803, "dur": 72.409,
    "args": {
      "Python parent id": 135, "Python id": 136, "finished": true, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875399.451, "dur": 0.380,
    "args": {
      "Python parent id": 136, "Python id": 137, "finished": true, "Ev Idx": 3037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875399.690, "dur": 0.053,
    "args": {
      "Python parent id": 137, "Python id": 138, "finished": true, "Ev Idx": 3038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875400.209, "dur": 0.385,
    "args": {
      "Python parent id": 136, "Python id": 139, "finished": true, "Ev Idx": 3039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875400.490, "dur": 0.055,
    "args": {
      "Python parent id": 139, "Python id": 140, "finished": true, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875400.959, "dur": 70.110,
    "args": {
      "Python parent id": 136, "Python id": 141, "finished": true, "Ev Idx": 3041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875401.704, "dur": 67.951,
    "args": {
      "Python parent id": 141, "Python id": 142, "finished": true, "Ev Idx": 3042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875409.004, "dur": 16.509,
    "args": {
      "Python parent id": 142, "Python id": 143, "finished": true, "Ev Idx": 3043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875409.144, "dur": 0.096,
    "args": {
      "Python parent id": 143, "Python id": 144, "finished": true, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875409.645, "dur": 15.738,
    "args": {
      "Python parent id": 143, "Python id": 145, "finished": true, "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875409.792, "dur": 7.398,
    "args": {
      "Python parent id": 145, "Python id": 146, "finished": true, "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875432.633, "dur": 6.816,
    "args": {
      "Python parent id": 142, "Python id": 147, "finished": true, "Ev Idx": 3047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875447.564, "dur": 8.315,
    "args": {
      "Python parent id": 142, "Python id": 148, "finished": true, "Ev Idx": 3048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875474.130, "dur": 0.143,
    "args": {
      "Python parent id": 12, "Python id": 149, "finished": true, "Ev Idx": 3049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875474.677, "dur": 0.170,
    "args": {
      "Python parent id": 12, "Python id": 150, "finished": true, "Ev Idx": 3050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875475.147, "dur": 0.061,
    "args": {
      "Python parent id": 12, "Python id": 151, "finished": true, "Ev Idx": 3051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875475.465, "dur": 8.862,
    "args": {
      "Python parent id": 12, "Python id": 152, "finished": true, "Ev Idx": 3052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875485.587, "dur": 84.772,
    "args": {
      "Python parent id": 12, "Python id": 153, "finished": true, "Ev Idx": 3053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875486.020, "dur": 0.463,
    "args": {
      "Python parent id": 153, "Python id": 154, "finished": true, "Ev Idx": 3054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875486.363, "dur": 0.054,
    "args": {
      "Python parent id": 154, "Python id": 155, "finished": true, "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875486.881, "dur": 0.377,
    "args": {
      "Python parent id": 153, "Python id": 156, "finished": true, "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875487.576, "dur": 0.259,
    "args": {
      "Python parent id": 153, "Python id": 157, "finished": true, "Ev Idx": 3057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875488.327, "dur": 1.330,
    "args": {
      "Python parent id": 153, "Python id": 158, "finished": true, "Ev Idx": 3058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875490.318, "dur": 79.940,
    "args": {
      "Python parent id": 153, "Python id": 159, "finished": true, "Ev Idx": 3059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875491.352, "dur": 78.705,
    "args": {
      "Python parent id": 159, "Python id": 160, "finished": true, "Ev Idx": 3060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875491.655, "dur": 78.320,
    "args": {
      "Python parent id": 160, "Python id": 161, "finished": true, "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875496.650, "dur": 72.163,
    "args": {
      "Python parent id": 161, "Python id": 162, "finished": true, "Ev Idx": 3062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875497.019, "dur": 71.737,
    "args": {
      "Python parent id": 162, "Python id": 163, "finished": true, "Ev Idx": 3063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875497.620, "dur": 71.032,
    "args": {
      "Python parent id": 163, "Python id": 164, "finished": true, "Ev Idx": 3064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875498.355, "dur": 0.377,
    "args": {
      "Python parent id": 164, "Python id": 165, "finished": true, "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875498.597, "dur": 0.051,
    "args": {
      "Python parent id": 165, "Python id": 166, "finished": true, "Ev Idx": 3066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875499.079, "dur": 0.534,
    "args": {
      "Python parent id": 164, "Python id": 167, "finished": true, "Ev Idx": 3067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875499.345, "dur": 0.193,
    "args": {
      "Python parent id": 167, "Python id": 168, "finished": true, "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875500.016, "dur": 68.507,
    "args": {
      "Python parent id": 164, "Python id": 169, "finished": true, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875500.956, "dur": 66.105,
    "args": {
      "Python parent id": 169, "Python id": 170, "finished": true, "Ev Idx": 3070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875507.965, "dur": 15.403,
    "args": {
      "Python parent id": 170, "Python id": 171, "finished": true, "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875508.106, "dur": 0.099,
    "args": {
      "Python parent id": 171, "Python id": 172, "finished": true, "Ev Idx": 3072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875508.537, "dur": 14.720,
    "args": {
      "Python parent id": 171, "Python id": 173, "finished": true, "Ev Idx": 3073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875508.638, "dur": 6.617,
    "args": {
      "Python parent id": 173, "Python id": 174, "finished": true, "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875531.069, "dur": 6.885,
    "args": {
      "Python parent id": 170, "Python id": 175, "finished": true, "Ev Idx": 3075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875545.582, "dur": 7.652,
    "args": {
      "Python parent id": 170, "Python id": 176, "finished": true, "Ev Idx": 3076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875571.114, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 177, "finished": true, "Ev Idx": 3077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875571.656, "dur": 0.160,
    "args": {
      "Python parent id": 12, "Python id": 178, "finished": true, "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875572.156, "dur": 0.041,
    "args": {
      "Python parent id": 12, "Python id": 179, "finished": true, "Ev Idx": 3079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875572.384, "dur": 9.105,
    "args": {
      "Python parent id": 12, "Python id": 180, "finished": true, "Ev Idx": 3080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875582.594, "dur": 82.627,
    "args": {
      "Python parent id": 12, "Python id": 181, "finished": true, "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875582.913, "dur": 0.588,
    "args": {
      "Python parent id": 181, "Python id": 182, "finished": true, "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875583.281, "dur": 0.157,
    "args": {
      "Python parent id": 182, "Python id": 183, "finished": true, "Ev Idx": 3083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875583.849, "dur": 0.389,
    "args": {
      "Python parent id": 181, "Python id": 184, "finished": true, "Ev Idx": 3084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875584.503, "dur": 0.261,
    "args": {
      "Python parent id": 181, "Python id": 185, "finished": true, "Ev Idx": 3085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875585.024, "dur": 1.253,
    "args": {
      "Python parent id": 181, "Python id": 186, "finished": true, "Ev Idx": 3086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875586.992, "dur": 78.133,
    "args": {
      "Python parent id": 181, "Python id": 187, "finished": true, "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875587.965, "dur": 76.968,
    "args": {
      "Python parent id": 187, "Python id": 188, "finished": true, "Ev Idx": 3088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875588.290, "dur": 76.588,
    "args": {
      "Python parent id": 188, "Python id": 189, "finished": true, "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875592.975, "dur": 70.843,
    "args": {
      "Python parent id": 189, "Python id": 190, "finished": true, "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875593.261, "dur": 70.504,
    "args": {
      "Python parent id": 190, "Python id": 191, "finished": true, "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875593.829, "dur": 69.790,
    "args": {
      "Python parent id": 191, "Python id": 192, "finished": true, "Ev Idx": 3092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875594.299, "dur": 0.455,
    "args": {
      "Python parent id": 192, "Python id": 193, "finished": true, "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875594.565, "dur": 0.119,
    "args": {
      "Python parent id": 193, "Python id": 194, "finished": true, "Ev Idx": 3094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875595.080, "dur": 0.332,
    "args": {
      "Python parent id": 192, "Python id": 195, "finished": true, "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875595.305, "dur": 0.058,
    "args": {
      "Python parent id": 195, "Python id": 196, "finished": true, "Ev Idx": 3096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875595.772, "dur": 67.742,
    "args": {
      "Python parent id": 192, "Python id": 197, "finished": true, "Ev Idx": 3097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875596.495, "dur": 65.708,
    "args": {
      "Python parent id": 197, "Python id": 198, "finished": true, "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875603.173, "dur": 15.578,
    "args": {
      "Python parent id": 198, "Python id": 199, "finished": true, "Ev Idx": 3099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875603.312, "dur": 0.092,
    "args": {
      "Python parent id": 199, "Python id": 200, "finished": true, "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875603.682, "dur": 14.985,
    "args": {
      "Python parent id": 199, "Python id": 201, "finished": true, "Ev Idx": 3101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875603.793, "dur": 6.956,
    "args": {
      "Python parent id": 201, "Python id": 202, "finished": true, "Ev Idx": 3102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875625.682, "dur": 6.547,
    "args": {
      "Python parent id": 198, "Python id": 203, "finished": true, "Ev Idx": 3103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875640.074, "dur": 8.003,
    "args": {
      "Python parent id": 198, "Python id": 204, "finished": true, "Ev Idx": 3104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875665.942, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 205, "finished": true, "Ev Idx": 3105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875666.353, "dur": 0.165,
    "args": {
      "Python parent id": 12, "Python id": 206, "finished": true, "Ev Idx": 3106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875666.743, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 207, "finished": true, "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875666.949, "dur": 8.805,
    "args": {
      "Python parent id": 12, "Python id": 208, "finished": true, "Ev Idx": 3108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875676.870, "dur": 97.211,
    "args": {
      "Python parent id": 12, "Python id": 209, "finished": true, "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875677.214, "dur": 0.511,
    "args": {
      "Python parent id": 209, "Python id": 210, "finished": true, "Ev Idx": 3110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875677.566, "dur": 0.113,
    "args": {
      "Python parent id": 210, "Python id": 211, "finished": true, "Ev Idx": 3111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875678.095, "dur": 0.405,
    "args": {
      "Python parent id": 209, "Python id": 212, "finished": true, "Ev Idx": 3112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875678.730, "dur": 0.256,
    "args": {
      "Python parent id": 209, "Python id": 213, "finished": true, "Ev Idx": 3113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875679.232, "dur": 1.227,
    "args": {
      "Python parent id": 209, "Python id": 214, "finished": true, "Ev Idx": 3114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875681.057, "dur": 92.954,
    "args": {
      "Python parent id": 209, "Python id": 215, "finished": true, "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875681.916, "dur": 91.936,
    "args": {
      "Python parent id": 215, "Python id": 216, "finished": true, "Ev Idx": 3116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875682.222, "dur": 91.553,
    "args": {
      "Python parent id": 216, "Python id": 217, "finished": true, "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875686.807, "dur": 85.965,
    "args": {
      "Python parent id": 217, "Python id": 218, "finished": true, "Ev Idx": 3118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875687.104, "dur": 85.614,
    "args": {
      "Python parent id": 218, "Python id": 219, "finished": true, "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875687.670, "dur": 84.899,
    "args": {
      "Python parent id": 219, "Python id": 220, "finished": true, "Ev Idx": 3120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875688.098, "dur": 0.365,
    "args": {
      "Python parent id": 220, "Python id": 221, "finished": true, "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875688.355, "dur": 0.048,
    "args": {
      "Python parent id": 221, "Python id": 222, "finished": true, "Ev Idx": 3122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875688.797, "dur": 0.318,
    "args": {
      "Python parent id": 220, "Python id": 223, "finished": true, "Ev Idx": 3123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875689.021, "dur": 0.040,
    "args": {
      "Python parent id": 223, "Python id": 224, "finished": true, "Ev Idx": 3124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875689.468, "dur": 83.020,
    "args": {
      "Python parent id": 220, "Python id": 225, "finished": true, "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875690.069, "dur": 81.220,
    "args": {
      "Python parent id": 225, "Python id": 226, "finished": true, "Ev Idx": 3126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875696.804, "dur": 30.392,
    "args": {
      "Python parent id": 226, "Python id": 227, "finished": true, "Ev Idx": 3127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875696.948, "dur": 0.116,
    "args": {
      "Python parent id": 227, "Python id": 228, "finished": true, "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875697.304, "dur": 29.821,
    "args": {
      "Python parent id": 227, "Python id": 229, "finished": true, "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875697.403, "dur": 6.957,
    "args": {
      "Python parent id": 229, "Python id": 230, "finished": true, "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875734.359, "dur": 6.561,
    "args": {
      "Python parent id": 226, "Python id": 231, "finished": true, "Ev Idx": 3131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875749.169, "dur": 7.671,
    "args": {
      "Python parent id": 226, "Python id": 232, "finished": true, "Ev Idx": 3132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875774.810, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 233, "finished": true, "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875775.235, "dur": 0.165,
    "args": {
      "Python parent id": 12, "Python id": 234, "finished": true, "Ev Idx": 3134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875775.599, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 235, "finished": true, "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875775.800, "dur": 30.436,
    "args": {
      "Python parent id": 12, "Python id": 236, "finished": true, "Ev Idx": 3136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875807.279, "dur": 84.600,
    "args": {
      "Python parent id": 12, "Python id": 237, "finished": true, "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875807.617, "dur": 0.519,
    "args": {
      "Python parent id": 237, "Python id": 238, "finished": true, "Ev Idx": 3138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875807.984, "dur": 0.095,
    "args": {
      "Python parent id": 238, "Python id": 239, "finished": true, "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875808.468, "dur": 0.433,
    "args": {
      "Python parent id": 237, "Python id": 240, "finished": true, "Ev Idx": 3140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875809.103, "dur": 0.264,
    "args": {
      "Python parent id": 237, "Python id": 241, "finished": true, "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875809.613, "dur": 1.269,
    "args": {
      "Python parent id": 237, "Python id": 242, "finished": true, "Ev Idx": 3142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875811.527, "dur": 80.281,
    "args": {
      "Python parent id": 237, "Python id": 243, "finished": true, "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875812.392, "dur": 79.258,
    "args": {
      "Python parent id": 243, "Python id": 244, "finished": true, "Ev Idx": 3144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875812.628, "dur": 78.947,
    "args": {
      "Python parent id": 244, "Python id": 245, "finished": true, "Ev Idx": 3145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875817.495, "dur": 72.935,
    "args": {
      "Python parent id": 245, "Python id": 246, "finished": true, "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875817.764, "dur": 72.597,
    "args": {
      "Python parent id": 246, "Python id": 247, "finished": true, "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875818.305, "dur": 71.925,
    "args": {
      "Python parent id": 247, "Python id": 248, "finished": true, "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875818.799, "dur": 0.367,
    "args": {
      "Python parent id": 248, "Python id": 249, "finished": true, "Ev Idx": 3149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875819.046, "dur": 0.067,
    "args": {
      "Python parent id": 249, "Python id": 250, "finished": true, "Ev Idx": 3150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875819.478, "dur": 0.357,
    "args": {
      "Python parent id": 248, "Python id": 251, "finished": true, "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875819.740, "dur": 0.041,
    "args": {
      "Python parent id": 251, "Python id": 252, "finished": true, "Ev Idx": 3152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875820.175, "dur": 69.895,
    "args": {
      "Python parent id": 248, "Python id": 253, "finished": true, "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875820.835, "dur": 68.001,
    "args": {
      "Python parent id": 253, "Python id": 254, "finished": true, "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875827.902, "dur": 16.715,
    "args": {
      "Python parent id": 254, "Python id": 255, "finished": true, "Ev Idx": 3155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875828.042, "dur": 0.093,
    "args": {
      "Python parent id": 255, "Python id": 256, "finished": true, "Ev Idx": 3156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875828.444, "dur": 16.068,
    "args": {
      "Python parent id": 255, "Python id": 257, "finished": true, "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875828.565, "dur": 7.374,
    "args": {
      "Python parent id": 257, "Python id": 258, "finished": true, "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875851.836, "dur": 6.537,
    "args": {
      "Python parent id": 254, "Python id": 259, "finished": true, "Ev Idx": 3159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875866.327, "dur": 7.799,
    "args": {
      "Python parent id": 254, "Python id": 260, "finished": true, "Ev Idx": 3160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875892.657, "dur": 0.109,
    "args": {
      "Python parent id": 12, "Python id": 261, "finished": true, "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875893.090, "dur": 0.176,
    "args": {
      "Python parent id": 12, "Python id": 262, "finished": true, "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875893.512, "dur": 0.046,
    "args": {
      "Python parent id": 12, "Python id": 263, "finished": true, "Ev Idx": 3163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875893.716, "dur": 9.074,
    "args": {
      "Python parent id": 12, "Python id": 264, "finished": true, "Ev Idx": 3164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875903.817, "dur": 83.712,
    "args": {
      "Python parent id": 12, "Python id": 265, "finished": true, "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875904.132, "dur": 0.508,
    "args": {
      "Python parent id": 265, "Python id": 266, "finished": true, "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875904.499, "dur": 0.076,
    "args": {
      "Python parent id": 266, "Python id": 267, "finished": true, "Ev Idx": 3167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875904.986, "dur": 0.327,
    "args": {
      "Python parent id": 265, "Python id": 268, "finished": true, "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875905.552, "dur": 0.226,
    "args": {
      "Python parent id": 265, "Python id": 269, "finished": true, "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740875906.003, "dur": 1.220,
    "args": {
      "Python parent id": 265, "Python id": 270, "finished": true, "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740875907.797, "dur": 79.626,
    "args": {
      "Python parent id": 265, "Python id": 271, "finished": true, "Ev Idx": 3171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740875908.471, "dur": 78.803,
    "args": {
      "Python parent id": 271, "Python id": 272, "finished": true, "Ev Idx": 3172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875908.702, "dur": 78.516,
    "args": {
      "Python parent id": 272, "Python id": 273, "finished": true, "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740875913.276, "dur": 72.834,
    "args": {
      "Python parent id": 273, "Python id": 274, "finished": true, "Ev Idx": 3174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740875913.592, "dur": 72.410,
    "args": {
      "Python parent id": 274, "Python id": 275, "finished": true, "Ev Idx": 3175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740875914.150, "dur": 71.756,
    "args": {
      "Python parent id": 275, "Python id": 276, "finished": true, "Ev Idx": 3176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875914.587, "dur": 0.420,
    "args": {
      "Python parent id": 276, "Python id": 277, "finished": true, "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875914.886, "dur": 0.074,
    "args": {
      "Python parent id": 277, "Python id": 278, "finished": true, "Ev Idx": 3178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875915.269, "dur": 0.359,
    "args": {
      "Python parent id": 276, "Python id": 279, "finished": true, "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875915.529, "dur": 0.045,
    "args": {
      "Python parent id": 279, "Python id": 280, "finished": true, "Ev Idx": 3180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875916.012, "dur": 69.783,
    "args": {
      "Python parent id": 276, "Python id": 281, "finished": true, "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740875916.552, "dur": 67.983,
    "args": {
      "Python parent id": 281, "Python id": 282, "finished": true, "Ev Idx": 3182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740875923.241, "dur": 16.810,
    "args": {
      "Python parent id": 282, "Python id": 283, "finished": true, "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740875923.360, "dur": 0.107,
    "args": {
      "Python parent id": 283, "Python id": 284, "finished": true, "Ev Idx": 3184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740875923.712, "dur": 16.199,
    "args": {
      "Python parent id": 283, "Python id": 285, "finished": true, "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740875923.825, "dur": 7.500,
    "args": {
      "Python parent id": 285, "Python id": 286, "finished": true, "Ev Idx": 3186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875947.073, "dur": 7.633,
    "args": {
      "Python parent id": 282, "Python id": 287, "finished": true, "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740875963.223, "dur": 7.538,
    "args": {
      "Python parent id": 282, "Python id": 288, "finished": true, "Ev Idx": 3188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740875988.177, "dur": 0.109,
    "args": {
      "Python parent id": 12, "Python id": 289, "finished": true, "Ev Idx": 3189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875988.605, "dur": 0.170,
    "args": {
      "Python parent id": 12, "Python id": 290, "finished": true, "Ev Idx": 3190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740875988.998, "dur": 0.043,
    "args": {
      "Python parent id": 12, "Python id": 291, "finished": true, "Ev Idx": 3191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740875989.203, "dur": 8.724,
    "args": {
      "Python parent id": 12, "Python id": 292, "finished": true, "Ev Idx": 3192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740875998.881, "dur": 83.175,
    "args": {
      "Python parent id": 12, "Python id": 293, "finished": true, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740875999.232, "dur": 0.498,
    "args": {
      "Python parent id": 293, "Python id": 294, "finished": true, "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740875999.569, "dur": 0.082,
    "args": {
      "Python parent id": 294, "Python id": 295, "finished": true, "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740875999.977, "dur": 0.377,
    "args": {
      "Python parent id": 293, "Python id": 296, "finished": true, "Ev Idx": 3196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876000.625, "dur": 0.257,
    "args": {
      "Python parent id": 293, "Python id": 297, "finished": true, "Ev Idx": 3197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876001.093, "dur": 1.208,
    "args": {
      "Python parent id": 293, "Python id": 298, "finished": true, "Ev Idx": 3198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876003.010, "dur": 78.941,
    "args": {
      "Python parent id": 293, "Python id": 299, "finished": true, "Ev Idx": 3199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876003.700, "dur": 78.108,
    "args": {
      "Python parent id": 299, "Python id": 300, "finished": true, "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876003.945, "dur": 77.785,
    "args": {
      "Python parent id": 300, "Python id": 301, "finished": true, "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876008.182, "dur": 72.328,
    "args": {
      "Python parent id": 301, "Python id": 302, "finished": true, "Ev Idx": 3202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876008.459, "dur": 71.984,
    "args": {
      "Python parent id": 302, "Python id": 303, "finished": true, "Ev Idx": 3203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876008.999, "dur": 71.338,
    "args": {
      "Python parent id": 303, "Python id": 304, "finished": true, "Ev Idx": 3204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876009.442, "dur": 0.395,
    "args": {
      "Python parent id": 304, "Python id": 305, "finished": true, "Ev Idx": 3205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876009.734, "dur": 0.045,
    "args": {
      "Python parent id": 305, "Python id": 306, "finished": true, "Ev Idx": 3206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876010.130, "dur": 0.316,
    "args": {
      "Python parent id": 304, "Python id": 307, "finished": true, "Ev Idx": 3207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876010.343, "dur": 0.039,
    "args": {
      "Python parent id": 307, "Python id": 308, "finished": true, "Ev Idx": 3208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876010.762, "dur": 69.496,
    "args": {
      "Python parent id": 304, "Python id": 309, "finished": true, "Ev Idx": 3209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876011.335, "dur": 67.725,
    "args": {
      "Python parent id": 309, "Python id": 310, "finished": true, "Ev Idx": 3210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876017.836, "dur": 16.729,
    "args": {
      "Python parent id": 310, "Python id": 311, "finished": true, "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876017.938, "dur": 0.098,
    "args": {
      "Python parent id": 311, "Python id": 312, "finished": true, "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876018.330, "dur": 16.157,
    "args": {
      "Python parent id": 311, "Python id": 313, "finished": true, "Ev Idx": 3213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876018.492, "dur": 7.563,
    "args": {
      "Python parent id": 313, "Python id": 314, "finished": true, "Ev Idx": 3214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876041.479, "dur": 6.378,
    "args": {
      "Python parent id": 310, "Python id": 315, "finished": true, "Ev Idx": 3215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876055.537, "dur": 7.571,
    "args": {
      "Python parent id": 310, "Python id": 316, "finished": true, "Ev Idx": 3216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876082.770, "dur": 0.119,
    "args": {
      "Python parent id": 12, "Python id": 317, "finished": true, "Ev Idx": 3217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876083.279, "dur": 0.184,
    "args": {
      "Python parent id": 12, "Python id": 318, "finished": true, "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876083.724, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 319, "finished": true, "Ev Idx": 3219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876083.928, "dur": 11.887,
    "args": {
      "Python parent id": 12, "Python id": 320, "finished": true, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876096.839, "dur": 86.251,
    "args": {
      "Python parent id": 12, "Python id": 321, "finished": true, "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876097.108, "dur": 0.436,
    "args": {
      "Python parent id": 321, "Python id": 322, "finished": true, "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876097.404, "dur": 0.099,
    "args": {
      "Python parent id": 322, "Python id": 323, "finished": true, "Ev Idx": 3223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876097.844, "dur": 0.360,
    "args": {
      "Python parent id": 321, "Python id": 324, "finished": true, "Ev Idx": 3224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876098.455, "dur": 0.245,
    "args": {
      "Python parent id": 321, "Python id": 325, "finished": true, "Ev Idx": 3225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876098.933, "dur": 1.364,
    "args": {
      "Python parent id": 321, "Python id": 326, "finished": true, "Ev Idx": 3226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876100.873, "dur": 82.117,
    "args": {
      "Python parent id": 321, "Python id": 327, "finished": true, "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876101.569, "dur": 81.271,
    "args": {
      "Python parent id": 327, "Python id": 328, "finished": true, "Ev Idx": 3228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876101.778, "dur": 81.007,
    "args": {
      "Python parent id": 328, "Python id": 329, "finished": true, "Ev Idx": 3229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876106.056, "dur": 75.478,
    "args": {
      "Python parent id": 329, "Python id": 330, "finished": true, "Ev Idx": 3230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876106.373, "dur": 75.101,
    "args": {
      "Python parent id": 330, "Python id": 331, "finished": true, "Ev Idx": 3231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876106.929, "dur": 74.417,
    "args": {
      "Python parent id": 331, "Python id": 332, "finished": true, "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876107.371, "dur": 0.395,
    "args": {
      "Python parent id": 332, "Python id": 333, "finished": true, "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876107.667, "dur": 0.041,
    "args": {
      "Python parent id": 333, "Python id": 334, "finished": true, "Ev Idx": 3234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876108.047, "dur": 0.346,
    "args": {
      "Python parent id": 332, "Python id": 335, "finished": true, "Ev Idx": 3235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876108.291, "dur": 0.045,
    "args": {
      "Python parent id": 335, "Python id": 336, "finished": true, "Ev Idx": 3236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876108.738, "dur": 72.530,
    "args": {
      "Python parent id": 332, "Python id": 337, "finished": true, "Ev Idx": 3237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876109.323, "dur": 70.752,
    "args": {
      "Python parent id": 337, "Python id": 338, "finished": true, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876115.905, "dur": 16.757,
    "args": {
      "Python parent id": 338, "Python id": 339, "finished": true, "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876116.030, "dur": 0.089,
    "args": {
      "Python parent id": 339, "Python id": 340, "finished": true, "Ev Idx": 3240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876116.387, "dur": 16.180,
    "args": {
      "Python parent id": 339, "Python id": 341, "finished": true, "Ev Idx": 3241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876116.533, "dur": 7.820,
    "args": {
      "Python parent id": 341, "Python id": 342, "finished": true, "Ev Idx": 3242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876140.276, "dur": 7.279,
    "args": {
      "Python parent id": 338, "Python id": 343, "finished": true, "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876156.686, "dur": 8.375,
    "args": {
      "Python parent id": 338, "Python id": 344, "finished": true, "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876183.732, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 345, "finished": true, "Ev Idx": 3245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876184.121, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 346, "finished": true, "Ev Idx": 3246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876184.542, "dur": 0.043,
    "args": {
      "Python parent id": 12, "Python id": 347, "finished": true, "Ev Idx": 3247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876184.797, "dur": 8.831,
    "args": {
      "Python parent id": 12, "Python id": 348, "finished": true, "Ev Idx": 3248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876194.727, "dur": 91.880,
    "args": {
      "Python parent id": 12, "Python id": 349, "finished": true, "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876195.113, "dur": 0.470,
    "args": {
      "Python parent id": 349, "Python id": 350, "finished": true, "Ev Idx": 3250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876195.449, "dur": 0.073,
    "args": {
      "Python parent id": 350, "Python id": 351, "finished": true, "Ev Idx": 3251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876196.006, "dur": 0.391,
    "args": {
      "Python parent id": 349, "Python id": 352, "finished": true, "Ev Idx": 3252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876196.678, "dur": 0.705,
    "args": {
      "Python parent id": 349, "Python id": 353, "finished": true, "Ev Idx": 3253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876197.725, "dur": 1.356,
    "args": {
      "Python parent id": 349, "Python id": 354, "finished": true, "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876199.683, "dur": 86.818,
    "args": {
      "Python parent id": 349, "Python id": 355, "finished": true, "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876200.604, "dur": 85.739,
    "args": {
      "Python parent id": 355, "Python id": 356, "finished": true, "Ev Idx": 3256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876200.861, "dur": 85.428,
    "args": {
      "Python parent id": 356, "Python id": 357, "finished": true, "Ev Idx": 3257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876205.326, "dur": 79.843,
    "args": {
      "Python parent id": 357, "Python id": 358, "finished": true, "Ev Idx": 3258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876205.568, "dur": 79.525,
    "args": {
      "Python parent id": 358, "Python id": 359, "finished": true, "Ev Idx": 3259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876206.114, "dur": 78.824,
    "args": {
      "Python parent id": 359, "Python id": 360, "finished": true, "Ev Idx": 3260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876206.734, "dur": 0.363,
    "args": {
      "Python parent id": 360, "Python id": 361, "finished": true, "Ev Idx": 3261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876206.992, "dur": 0.048,
    "args": {
      "Python parent id": 361, "Python id": 362, "finished": true, "Ev Idx": 3262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876207.474, "dur": 0.310,
    "args": {
      "Python parent id": 360, "Python id": 363, "finished": true, "Ev Idx": 3263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876207.685, "dur": 0.045,
    "args": {
      "Python parent id": 363, "Python id": 364, "finished": true, "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876208.096, "dur": 76.738,
    "args": {
      "Python parent id": 360, "Python id": 365, "finished": true, "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876208.952, "dur": 74.499,
    "args": {
      "Python parent id": 365, "Python id": 366, "finished": true, "Ev Idx": 3266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876216.086, "dur": 16.410,
    "args": {
      "Python parent id": 366, "Python id": 367, "finished": true, "Ev Idx": 3267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876216.176, "dur": 0.088,
    "args": {
      "Python parent id": 367, "Python id": 368, "finished": true, "Ev Idx": 3268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876216.523, "dur": 15.857,
    "args": {
      "Python parent id": 367, "Python id": 369, "finished": true, "Ev Idx": 3269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876216.645, "dur": 6.754,
    "args": {
      "Python parent id": 369, "Python id": 370, "finished": true, "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876241.561, "dur": 8.686,
    "args": {
      "Python parent id": 366, "Python id": 371, "finished": true, "Ev Idx": 3271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876259.581, "dur": 8.589,
    "args": {
      "Python parent id": 366, "Python id": 372, "finished": true, "Ev Idx": 3272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876287.307, "dur": 0.083,
    "args": {
      "Python parent id": 12, "Python id": 373, "finished": true, "Ev Idx": 3273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876287.732, "dur": 0.159,
    "args": {
      "Python parent id": 12, "Python id": 374, "finished": true, "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876288.123, "dur": 0.048,
    "args": {
      "Python parent id": 12, "Python id": 375, "finished": true, "Ev Idx": 3275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876288.355, "dur": 10.108,
    "args": {
      "Python parent id": 12, "Python id": 376, "finished": true, "Ev Idx": 3276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876299.514, "dur": 89.016,
    "args": {
      "Python parent id": 12, "Python id": 377, "finished": true, "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876299.852, "dur": 0.538,
    "args": {
      "Python parent id": 377, "Python id": 378, "finished": true, "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876300.213, "dur": 0.085,
    "args": {
      "Python parent id": 378, "Python id": 379, "finished": true, "Ev Idx": 3279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876300.689, "dur": 0.372,
    "args": {
      "Python parent id": 377, "Python id": 380, "finished": true, "Ev Idx": 3280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876301.406, "dur": 0.251,
    "args": {
      "Python parent id": 377, "Python id": 381, "finished": true, "Ev Idx": 3281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876301.934, "dur": 1.222,
    "args": {
      "Python parent id": 377, "Python id": 382, "finished": true, "Ev Idx": 3282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876303.731, "dur": 84.719,
    "args": {
      "Python parent id": 377, "Python id": 383, "finished": true, "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876304.518, "dur": 83.776,
    "args": {
      "Python parent id": 383, "Python id": 384, "finished": true, "Ev Idx": 3284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876304.707, "dur": 83.530,
    "args": {
      "Python parent id": 384, "Python id": 385, "finished": true, "Ev Idx": 3285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876308.989, "dur": 78.177,
    "args": {
      "Python parent id": 385, "Python id": 386, "finished": true, "Ev Idx": 3286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876309.293, "dur": 77.766,
    "args": {
      "Python parent id": 386, "Python id": 387, "finished": true, "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876309.819, "dur": 77.161,
    "args": {
      "Python parent id": 387, "Python id": 388, "finished": true, "Ev Idx": 3288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876310.325, "dur": 0.400,
    "args": {
      "Python parent id": 388, "Python id": 389, "finished": true, "Ev Idx": 3289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876310.605, "dur": 0.068,
    "args": {
      "Python parent id": 389, "Python id": 390, "finished": true, "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876311.012, "dur": 0.327,
    "args": {
      "Python parent id": 388, "Python id": 391, "finished": true, "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876311.225, "dur": 0.071,
    "args": {
      "Python parent id": 391, "Python id": 392, "finished": true, "Ev Idx": 3292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876311.704, "dur": 75.181,
    "args": {
      "Python parent id": 388, "Python id": 393, "finished": true, "Ev Idx": 3293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876312.317, "dur": 73.338,
    "args": {
      "Python parent id": 393, "Python id": 394, "finished": true, "Ev Idx": 3294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876319.213, "dur": 17.494,
    "args": {
      "Python parent id": 394, "Python id": 395, "finished": true, "Ev Idx": 3295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876319.323, "dur": 0.093,
    "args": {
      "Python parent id": 395, "Python id": 396, "finished": true, "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876319.681, "dur": 16.923,
    "args": {
      "Python parent id": 395, "Python id": 397, "finished": true, "Ev Idx": 3297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876319.822, "dur": 7.859,
    "args": {
      "Python parent id": 397, "Python id": 398, "finished": true, "Ev Idx": 3298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876344.619, "dur": 6.932,
    "args": {
      "Python parent id": 394, "Python id": 399, "finished": true, "Ev Idx": 3299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876361.325, "dur": 8.369,
    "args": {
      "Python parent id": 394, "Python id": 400, "finished": true, "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876389.184, "dur": 0.107,
    "args": {
      "Python parent id": 12, "Python id": 401, "finished": true, "Ev Idx": 3301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876389.647, "dur": 0.204,
    "args": {
      "Python parent id": 12, "Python id": 402, "finished": true, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876390.127, "dur": 0.040,
    "args": {
      "Python parent id": 12, "Python id": 403, "finished": true, "Ev Idx": 3303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876390.361, "dur": 9.357,
    "args": {
      "Python parent id": 12, "Python id": 404, "finished": true, "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876400.744, "dur": 87.106,
    "args": {
      "Python parent id": 12, "Python id": 405, "finished": true, "Ev Idx": 3305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876401.086, "dur": 0.463,
    "args": {
      "Python parent id": 405, "Python id": 406, "finished": true, "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876401.404, "dur": 0.084,
    "args": {
      "Python parent id": 406, "Python id": 407, "finished": true, "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876401.829, "dur": 0.364,
    "args": {
      "Python parent id": 405, "Python id": 408, "finished": true, "Ev Idx": 3308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876402.428, "dur": 0.232,
    "args": {
      "Python parent id": 405, "Python id": 409, "finished": true, "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876402.916, "dur": 1.203,
    "args": {
      "Python parent id": 405, "Python id": 410, "finished": true, "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876404.611, "dur": 83.158,
    "args": {
      "Python parent id": 405, "Python id": 411, "finished": true, "Ev Idx": 3311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876405.236, "dur": 82.390,
    "args": {
      "Python parent id": 411, "Python id": 412, "finished": true, "Ev Idx": 3312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876405.465, "dur": 82.104,
    "args": {
      "Python parent id": 412, "Python id": 413, "finished": true, "Ev Idx": 3313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876409.953, "dur": 76.525,
    "args": {
      "Python parent id": 413, "Python id": 414, "finished": true, "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876410.261, "dur": 76.131,
    "args": {
      "Python parent id": 414, "Python id": 415, "finished": true, "Ev Idx": 3315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876410.743, "dur": 75.576,
    "args": {
      "Python parent id": 415, "Python id": 416, "finished": true, "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876411.203, "dur": 0.352,
    "args": {
      "Python parent id": 416, "Python id": 417, "finished": true, "Ev Idx": 3317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876411.433, "dur": 0.069,
    "args": {
      "Python parent id": 417, "Python id": 418, "finished": true, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876411.802, "dur": 0.324,
    "args": {
      "Python parent id": 416, "Python id": 419, "finished": true, "Ev Idx": 3319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876412.032, "dur": 0.039,
    "args": {
      "Python parent id": 419, "Python id": 420, "finished": true, "Ev Idx": 3320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876412.478, "dur": 73.722,
    "args": {
      "Python parent id": 416, "Python id": 421, "finished": true, "Ev Idx": 3321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876413.015, "dur": 71.986,
    "args": {
      "Python parent id": 421, "Python id": 422, "finished": true, "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876419.873, "dur": 16.934,
    "args": {
      "Python parent id": 422, "Python id": 423, "finished": true, "Ev Idx": 3323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876420.003, "dur": 0.109,
    "args": {
      "Python parent id": 423, "Python id": 424, "finished": true, "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876420.379, "dur": 16.335,
    "args": {
      "Python parent id": 423, "Python id": 425, "finished": true, "Ev Idx": 3325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876420.505, "dur": 7.291,
    "args": {
      "Python parent id": 425, "Python id": 426, "finished": true, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876445.974, "dur": 6.957,
    "args": {
      "Python parent id": 422, "Python id": 427, "finished": true, "Ev Idx": 3327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876461.513, "dur": 8.068,
    "args": {
      "Python parent id": 422, "Python id": 428, "finished": true, "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876488.474, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 429, "finished": true, "Ev Idx": 3329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876488.892, "dur": 0.160,
    "args": {
      "Python parent id": 12, "Python id": 430, "finished": true, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876489.260, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 431, "finished": true, "Ev Idx": 3331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876489.481, "dur": 8.594,
    "args": {
      "Python parent id": 12, "Python id": 432, "finished": true, "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876499.172, "dur": 84.339,
    "args": {
      "Python parent id": 12, "Python id": 433, "finished": true, "Ev Idx": 3333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876499.532, "dur": 0.436,
    "args": {
      "Python parent id": 433, "Python id": 434, "finished": true, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876499.835, "dur": 0.073,
    "args": {
      "Python parent id": 434, "Python id": 435, "finished": true, "Ev Idx": 3335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876500.234, "dur": 0.331,
    "args": {
      "Python parent id": 433, "Python id": 436, "finished": true, "Ev Idx": 3336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876500.782, "dur": 0.223,
    "args": {
      "Python parent id": 433, "Python id": 437, "finished": true, "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876501.228, "dur": 1.252,
    "args": {
      "Python parent id": 433, "Python id": 438, "finished": true, "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876503.021, "dur": 80.414,
    "args": {
      "Python parent id": 433, "Python id": 439, "finished": true, "Ev Idx": 3339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876503.792, "dur": 79.485,
    "args": {
      "Python parent id": 439, "Python id": 440, "finished": true, "Ev Idx": 3340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876503.983, "dur": 79.213,
    "args": {
      "Python parent id": 440, "Python id": 441, "finished": true, "Ev Idx": 3341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876508.341, "dur": 73.848,
    "args": {
      "Python parent id": 441, "Python id": 442, "finished": true, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876508.559, "dur": 73.496,
    "args": {
      "Python parent id": 442, "Python id": 443, "finished": true, "Ev Idx": 3343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876509.039, "dur": 72.936,
    "args": {
      "Python parent id": 443, "Python id": 444, "finished": true, "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876509.551, "dur": 0.342,
    "args": {
      "Python parent id": 444, "Python id": 445, "finished": true, "Ev Idx": 3345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876509.789, "dur": 0.045,
    "args": {
      "Python parent id": 445, "Python id": 446, "finished": true, "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876510.182, "dur": 0.335,
    "args": {
      "Python parent id": 444, "Python id": 447, "finished": true, "Ev Idx": 3347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876510.417, "dur": 0.044,
    "args": {
      "Python parent id": 447, "Python id": 448, "finished": true, "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876510.835, "dur": 71.054,
    "args": {
      "Python parent id": 444, "Python id": 449, "finished": true, "Ev Idx": 3349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876511.371, "dur": 69.271,
    "args": {
      "Python parent id": 449, "Python id": 450, "finished": true, "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876518.459, "dur": 17.123,
    "args": {
      "Python parent id": 450, "Python id": 451, "finished": true, "Ev Idx": 3351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876518.556, "dur": 0.093,
    "args": {
      "Python parent id": 451, "Python id": 452, "finished": true, "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876518.900, "dur": 16.561,
    "args": {
      "Python parent id": 451, "Python id": 453, "finished": true, "Ev Idx": 3353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876519.052, "dur": 7.765,
    "args": {
      "Python parent id": 453, "Python id": 454, "finished": true, "Ev Idx": 3354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876542.809, "dur": 7.256,
    "args": {
      "Python parent id": 450, "Python id": 455, "finished": true, "Ev Idx": 3355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876558.379, "dur": 7.414,
    "args": {
      "Python parent id": 450, "Python id": 456, "finished": true, "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876584.193, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 457, "finished": true, "Ev Idx": 3357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876584.589, "dur": 0.156,
    "args": {
      "Python parent id": 12, "Python id": 458, "finished": true, "Ev Idx": 3358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876584.953, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 459, "finished": true, "Ev Idx": 3359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876585.165, "dur": 9.744,
    "args": {
      "Python parent id": 12, "Python id": 460, "finished": true, "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876595.930, "dur": 87.028,
    "args": {
      "Python parent id": 12, "Python id": 461, "finished": true, "Ev Idx": 3361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876596.253, "dur": 0.469,
    "args": {
      "Python parent id": 461, "Python id": 462, "finished": true, "Ev Idx": 3362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876596.590, "dur": 0.074,
    "args": {
      "Python parent id": 462, "Python id": 463, "finished": true, "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876596.951, "dur": 0.324,
    "args": {
      "Python parent id": 461, "Python id": 464, "finished": true, "Ev Idx": 3364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876597.489, "dur": 0.260,
    "args": {
      "Python parent id": 461, "Python id": 465, "finished": true, "Ev Idx": 3365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876598.021, "dur": 1.213,
    "args": {
      "Python parent id": 461, "Python id": 466, "finished": true, "Ev Idx": 3366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876599.787, "dur": 83.097,
    "args": {
      "Python parent id": 461, "Python id": 467, "finished": true, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876600.469, "dur": 82.243,
    "args": {
      "Python parent id": 467, "Python id": 468, "finished": true, "Ev Idx": 3368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876600.686, "dur": 81.940,
    "args": {
      "Python parent id": 468, "Python id": 469, "finished": true, "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876605.045, "dur": 76.535,
    "args": {
      "Python parent id": 469, "Python id": 470, "finished": true, "Ev Idx": 3370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876605.333, "dur": 76.178,
    "args": {
      "Python parent id": 470, "Python id": 471, "finished": true, "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876605.799, "dur": 75.634,
    "args": {
      "Python parent id": 471, "Python id": 472, "finished": true, "Ev Idx": 3372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876606.274, "dur": 0.431,
    "args": {
      "Python parent id": 472, "Python id": 473, "finished": true, "Ev Idx": 3373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876606.582, "dur": 0.074,
    "args": {
      "Python parent id": 473, "Python id": 474, "finished": true, "Ev Idx": 3374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876606.976, "dur": 0.395,
    "args": {
      "Python parent id": 472, "Python id": 475, "finished": true, "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876607.284, "dur": 0.034,
    "args": {
      "Python parent id": 475, "Python id": 476, "finished": true, "Ev Idx": 3376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876607.742, "dur": 73.581,
    "args": {
      "Python parent id": 472, "Python id": 477, "finished": true, "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876608.374, "dur": 71.677,
    "args": {
      "Python parent id": 477, "Python id": 478, "finished": true, "Ev Idx": 3378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876614.892, "dur": 17.386,
    "args": {
      "Python parent id": 478, "Python id": 479, "finished": true, "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876615.046, "dur": 0.089,
    "args": {
      "Python parent id": 479, "Python id": 480, "finished": true, "Ev Idx": 3380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876615.445, "dur": 16.733,
    "args": {
      "Python parent id": 479, "Python id": 481, "finished": true, "Ev Idx": 3381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876615.590, "dur": 7.664,
    "args": {
      "Python parent id": 481, "Python id": 482, "finished": true, "Ev Idx": 3382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876639.966, "dur": 6.854,
    "args": {
      "Python parent id": 478, "Python id": 483, "finished": true, "Ev Idx": 3383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876656.068, "dur": 8.647,
    "args": {
      "Python parent id": 478, "Python id": 484, "finished": true, "Ev Idx": 3384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876683.643, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 485, "finished": true, "Ev Idx": 3385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876684.050, "dur": 0.161,
    "args": {
      "Python parent id": 12, "Python id": 486, "finished": true, "Ev Idx": 3386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876684.427, "dur": 0.046,
    "args": {
      "Python parent id": 12, "Python id": 487, "finished": true, "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876684.634, "dur": 9.516,
    "args": {
      "Python parent id": 12, "Python id": 488, "finished": true, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876695.235, "dur": 84.317,
    "args": {
      "Python parent id": 12, "Python id": 489, "finished": true, "Ev Idx": 3389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876695.559, "dur": 0.465,
    "args": {
      "Python parent id": 489, "Python id": 490, "finished": true, "Ev Idx": 3390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876695.875, "dur": 0.087,
    "args": {
      "Python parent id": 490, "Python id": 491, "finished": true, "Ev Idx": 3391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876696.272, "dur": 0.344,
    "args": {
      "Python parent id": 489, "Python id": 492, "finished": true, "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876696.816, "dur": 0.249,
    "args": {
      "Python parent id": 489, "Python id": 493, "finished": true, "Ev Idx": 3393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876697.290, "dur": 1.288,
    "args": {
      "Python parent id": 489, "Python id": 494, "finished": true, "Ev Idx": 3394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876699.126, "dur": 80.355,
    "args": {
      "Python parent id": 489, "Python id": 495, "finished": true, "Ev Idx": 3395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876699.783, "dur": 79.518,
    "args": {
      "Python parent id": 495, "Python id": 496, "finished": true, "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876699.991, "dur": 79.250,
    "args": {
      "Python parent id": 496, "Python id": 497, "finished": true, "Ev Idx": 3397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876704.299, "dur": 73.853,
    "args": {
      "Python parent id": 497, "Python id": 498, "finished": true, "Ev Idx": 3398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876704.584, "dur": 73.482,
    "args": {
      "Python parent id": 498, "Python id": 499, "finished": true, "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876705.098, "dur": 72.898,
    "args": {
      "Python parent id": 499, "Python id": 500, "finished": true, "Ev Idx": 3400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876705.586, "dur": 0.327,
    "args": {
      "Python parent id": 500, "Python id": 501, "finished": true, "Ev Idx": 3401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876705.811, "dur": 0.043,
    "args": {
      "Python parent id": 501, "Python id": 502, "finished": true, "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876706.163, "dur": 0.346,
    "args": {
      "Python parent id": 500, "Python id": 503, "finished": true, "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876706.420, "dur": 0.036,
    "args": {
      "Python parent id": 503, "Python id": 504, "finished": true, "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876706.855, "dur": 70.986,
    "args": {
      "Python parent id": 500, "Python id": 505, "finished": true, "Ev Idx": 3405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876707.417, "dur": 69.231,
    "args": {
      "Python parent id": 505, "Python id": 506, "finished": true, "Ev Idx": 3406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876714.119, "dur": 15.378,
    "args": {
      "Python parent id": 506, "Python id": 507, "finished": true, "Ev Idx": 3407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876714.218, "dur": 0.089,
    "args": {
      "Python parent id": 507, "Python id": 508, "finished": true, "Ev Idx": 3408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876714.587, "dur": 14.834,
    "args": {
      "Python parent id": 507, "Python id": 509, "finished": true, "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876714.690, "dur": 6.809,
    "args": {
      "Python parent id": 509, "Python id": 510, "finished": true, "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876736.195, "dur": 8.154,
    "args": {
      "Python parent id": 506, "Python id": 511, "finished": true, "Ev Idx": 3411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876752.794, "dur": 8.126,
    "args": {
      "Python parent id": 506, "Python id": 512, "finished": true, "Ev Idx": 3412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876780.148, "dur": 0.106,
    "args": {
      "Python parent id": 12, "Python id": 513, "finished": true, "Ev Idx": 3413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876780.535, "dur": 0.183,
    "args": {
      "Python parent id": 12, "Python id": 514, "finished": true, "Ev Idx": 3414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876780.980, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 515, "finished": true, "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876781.205, "dur": 8.736,
    "args": {
      "Python parent id": 12, "Python id": 516, "finished": true, "Ev Idx": 3416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876790.946, "dur": 88.065,
    "args": {
      "Python parent id": 12, "Python id": 517, "finished": true, "Ev Idx": 3417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876791.237, "dur": 0.471,
    "args": {
      "Python parent id": 517, "Python id": 518, "finished": true, "Ev Idx": 3418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876791.556, "dur": 0.085,
    "args": {
      "Python parent id": 518, "Python id": 519, "finished": true, "Ev Idx": 3419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876791.951, "dur": 0.344,
    "args": {
      "Python parent id": 517, "Python id": 520, "finished": true, "Ev Idx": 3420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876792.514, "dur": 0.208,
    "args": {
      "Python parent id": 517, "Python id": 521, "finished": true, "Ev Idx": 3421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876792.940, "dur": 1.226,
    "args": {
      "Python parent id": 517, "Python id": 522, "finished": true, "Ev Idx": 3422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876794.660, "dur": 84.254,
    "args": {
      "Python parent id": 517, "Python id": 523, "finished": true, "Ev Idx": 3423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876795.356, "dur": 83.374,
    "args": {
      "Python parent id": 523, "Python id": 524, "finished": true, "Ev Idx": 3424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876795.588, "dur": 83.063,
    "args": {
      "Python parent id": 524, "Python id": 525, "finished": true, "Ev Idx": 3425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876800.002, "dur": 77.653,
    "args": {
      "Python parent id": 525, "Python id": 526, "finished": true, "Ev Idx": 3426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876800.242, "dur": 77.325,
    "args": {
      "Python parent id": 526, "Python id": 527, "finished": true, "Ev Idx": 3427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876800.762, "dur": 76.669,
    "args": {
      "Python parent id": 527, "Python id": 528, "finished": true, "Ev Idx": 3428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876801.208, "dur": 1.408,
    "args": {
      "Python parent id": 528, "Python id": 529, "finished": true, "Ev Idx": 3429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876801.455, "dur": 0.871,
    "args": {
      "Python parent id": 529, "Python id": 530, "finished": true, "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876802.933, "dur": 0.379,
    "args": {
      "Python parent id": 528, "Python id": 531, "finished": true, "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876803.212, "dur": 0.040,
    "args": {
      "Python parent id": 531, "Python id": 532, "finished": true, "Ev Idx": 3432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876803.632, "dur": 73.679,
    "args": {
      "Python parent id": 528, "Python id": 533, "finished": true, "Ev Idx": 3433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876804.236, "dur": 71.916,
    "args": {
      "Python parent id": 533, "Python id": 534, "finished": true, "Ev Idx": 3434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876810.980, "dur": 17.566,
    "args": {
      "Python parent id": 534, "Python id": 535, "finished": true, "Ev Idx": 3435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876811.082, "dur": 0.090,
    "args": {
      "Python parent id": 535, "Python id": 536, "finished": true, "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876811.426, "dur": 17.050,
    "args": {
      "Python parent id": 535, "Python id": 537, "finished": true, "Ev Idx": 3437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876811.541, "dur": 8.108,
    "args": {
      "Python parent id": 537, "Python id": 538, "finished": true, "Ev Idx": 3438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876836.150, "dur": 7.396,
    "args": {
      "Python parent id": 534, "Python id": 539, "finished": true, "Ev Idx": 3439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876852.107, "dur": 7.773,
    "args": {
      "Python parent id": 534, "Python id": 540, "finished": true, "Ev Idx": 3440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740876879.674, "dur": 0.107,
    "args": {
      "Python parent id": 12, "Python id": 541, "finished": true, "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876880.090, "dur": 0.172,
    "args": {
      "Python parent id": 12, "Python id": 542, "finished": true, "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740876880.495, "dur": 0.059,
    "args": {
      "Python parent id": 12, "Python id": 543, "finished": true, "Ev Idx": 3443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740876880.730, "dur": 9.363,
    "args": {
      "Python parent id": 12, "Python id": 544, "finished": true, "Ev Idx": 3444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740876891.155, "dur": 116.952,
    "args": {
      "Python parent id": 12, "Python id": 545, "finished": true, "Ev Idx": 3445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876891.475, "dur": 0.439,
    "args": {
      "Python parent id": 545, "Python id": 546, "finished": true, "Ev Idx": 3446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876891.757, "dur": 0.086,
    "args": {
      "Python parent id": 546, "Python id": 547, "finished": true, "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876892.159, "dur": 0.343,
    "args": {
      "Python parent id": 545, "Python id": 548, "finished": true, "Ev Idx": 3448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740876892.720, "dur": 0.207,
    "args": {
      "Python parent id": 545, "Python id": 549, "finished": true, "Ev Idx": 3449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740876893.187, "dur": 1.254,
    "args": {
      "Python parent id": 545, "Python id": 550, "finished": true, "Ev Idx": 3450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740876895.037, "dur": 112.993,
    "args": {
      "Python parent id": 545, "Python id": 551, "finished": true, "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740876895.703, "dur": 112.165,
    "args": {
      "Python parent id": 551, "Python id": 552, "finished": true, "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876895.895, "dur": 111.890,
    "args": {
      "Python parent id": 552, "Python id": 553, "finished": true, "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740876900.036, "dur": 106.696,
    "args": {
      "Python parent id": 553, "Python id": 554, "finished": true, "Ev Idx": 3454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740876900.335, "dur": 106.320,
    "args": {
      "Python parent id": 554, "Python id": 555, "finished": true, "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740876900.881, "dur": 105.649,
    "args": {
      "Python parent id": 555, "Python id": 556, "finished": true, "Ev Idx": 3456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876901.332, "dur": 0.346,
    "args": {
      "Python parent id": 556, "Python id": 557, "finished": true, "Ev Idx": 3457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876901.573, "dur": 0.046,
    "args": {
      "Python parent id": 557, "Python id": 558, "finished": true, "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740876901.965, "dur": 0.310,
    "args": {
      "Python parent id": 556, "Python id": 559, "finished": true, "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740876902.176, "dur": 0.043,
    "args": {
      "Python parent id": 559, "Python id": 560, "finished": true, "Ev Idx": 3460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876902.602, "dur": 103.848,
    "args": {
      "Python parent id": 556, "Python id": 561, "finished": true, "Ev Idx": 3461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740876903.139, "dur": 101.974,
    "args": {
      "Python parent id": 561, "Python id": 562, "finished": true, "Ev Idx": 3462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740876909.621, "dur": 45.098,
    "args": {
      "Python parent id": 562, "Python id": 563, "finished": true, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740876909.722, "dur": 0.085,
    "args": {
      "Python parent id": 563, "Python id": 564, "finished": true, "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740876910.074, "dur": 44.547,
    "args": {
      "Python parent id": 563, "Python id": 565, "finished": true, "Ev Idx": 3465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740876910.179, "dur": 34.901,
    "args": {
      "Python parent id": 565, "Python id": 566, "finished": true, "Ev Idx": 3466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876963.184, "dur": 8.145,
    "args": {
      "Python parent id": 562, "Python id": 567, "finished": true, "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740876980.607, "dur": 8.739,
    "args": {
      "Python parent id": 562, "Python id": 568, "finished": true, "Ev Idx": 3468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877009.059, "dur": 0.132,
    "args": {
      "Python parent id": 12, "Python id": 569, "finished": true, "Ev Idx": 3469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877009.623, "dur": 0.183,
    "args": {
      "Python parent id": 12, "Python id": 570, "finished": true, "Ev Idx": 3470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877010.083, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 571, "finished": true, "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877010.310, "dur": 10.457,
    "args": {
      "Python parent id": 12, "Python id": 572, "finished": true, "Ev Idx": 3472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877021.777, "dur": 91.469,
    "args": {
      "Python parent id": 12, "Python id": 573, "finished": true, "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877022.120, "dur": 0.419,
    "args": {
      "Python parent id": 573, "Python id": 574, "finished": true, "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877022.394, "dur": 0.085,
    "args": {
      "Python parent id": 574, "Python id": 575, "finished": true, "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877022.803, "dur": 0.345,
    "args": {
      "Python parent id": 573, "Python id": 576, "finished": true, "Ev Idx": 3476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877023.352, "dur": 0.251,
    "args": {
      "Python parent id": 573, "Python id": 577, "finished": true, "Ev Idx": 3477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877023.885, "dur": 1.286,
    "args": {
      "Python parent id": 573, "Python id": 578, "finished": true, "Ev Idx": 3478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877025.733, "dur": 87.439,
    "args": {
      "Python parent id": 573, "Python id": 579, "finished": true, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877026.487, "dur": 86.538,
    "args": {
      "Python parent id": 579, "Python id": 580, "finished": true, "Ev Idx": 3480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877026.760, "dur": 86.206,
    "args": {
      "Python parent id": 580, "Python id": 581, "finished": true, "Ev Idx": 3481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877031.493, "dur": 80.446,
    "args": {
      "Python parent id": 581, "Python id": 582, "finished": true, "Ev Idx": 3482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877031.793, "dur": 79.985,
    "args": {
      "Python parent id": 582, "Python id": 583, "finished": true, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877032.334, "dur": 79.325,
    "args": {
      "Python parent id": 583, "Python id": 584, "finished": true, "Ev Idx": 3484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877032.875, "dur": 0.380,
    "args": {
      "Python parent id": 584, "Python id": 585, "finished": true, "Ev Idx": 3485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877033.139, "dur": 0.049,
    "args": {
      "Python parent id": 585, "Python id": 586, "finished": true, "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877033.552, "dur": 0.341,
    "args": {
      "Python parent id": 584, "Python id": 587, "finished": true, "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877033.792, "dur": 0.042,
    "args": {
      "Python parent id": 587, "Python id": 588, "finished": true, "Ev Idx": 3488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877034.191, "dur": 77.398,
    "args": {
      "Python parent id": 584, "Python id": 589, "finished": true, "Ev Idx": 3489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877034.744, "dur": 75.713,
    "args": {
      "Python parent id": 589, "Python id": 590, "finished": true, "Ev Idx": 3490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877041.435, "dur": 20.361,
    "args": {
      "Python parent id": 590, "Python id": 591, "finished": true, "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877041.569, "dur": 0.118,
    "args": {
      "Python parent id": 591, "Python id": 592, "finished": true, "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877041.999, "dur": 19.716,
    "args": {
      "Python parent id": 591, "Python id": 593, "finished": true, "Ev Idx": 3493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877042.109, "dur": 8.280,
    "args": {
      "Python parent id": 593, "Python id": 594, "finished": true, "Ev Idx": 3494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877069.052, "dur": 6.963,
    "args": {
      "Python parent id": 590, "Python id": 595, "finished": true, "Ev Idx": 3495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877084.846, "dur": 9.467,
    "args": {
      "Python parent id": 590, "Python id": 596, "finished": true, "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877113.881, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 597, "finished": true, "Ev Idx": 3497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877114.335, "dur": 0.153,
    "args": {
      "Python parent id": 12, "Python id": 598, "finished": true, "Ev Idx": 3498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877114.729, "dur": 0.048,
    "args": {
      "Python parent id": 12, "Python id": 599, "finished": true, "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877114.946, "dur": 10.242,
    "args": {
      "Python parent id": 12, "Python id": 600, "finished": true, "Ev Idx": 3500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877126.218, "dur": 88.472,
    "args": {
      "Python parent id": 12, "Python id": 601, "finished": true, "Ev Idx": 3501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877126.590, "dur": 0.385,
    "args": {
      "Python parent id": 601, "Python id": 602, "finished": true, "Ev Idx": 3502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877126.831, "dur": 0.086,
    "args": {
      "Python parent id": 602, "Python id": 603, "finished": true, "Ev Idx": 3503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877127.232, "dur": 0.354,
    "args": {
      "Python parent id": 601, "Python id": 604, "finished": true, "Ev Idx": 3504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877127.805, "dur": 0.239,
    "args": {
      "Python parent id": 601, "Python id": 605, "finished": true, "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877128.260, "dur": 1.259,
    "args": {
      "Python parent id": 601, "Python id": 606, "finished": true, "Ev Idx": 3506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877130.058, "dur": 84.558,
    "args": {
      "Python parent id": 601, "Python id": 607, "finished": true, "Ev Idx": 3507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877130.746, "dur": 83.696,
    "args": {
      "Python parent id": 607, "Python id": 608, "finished": true, "Ev Idx": 3508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877130.952, "dur": 83.432,
    "args": {
      "Python parent id": 608, "Python id": 609, "finished": true, "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877135.677, "dur": 77.747,
    "args": {
      "Python parent id": 609, "Python id": 610, "finished": true, "Ev Idx": 3510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877135.956, "dur": 77.383,
    "args": {
      "Python parent id": 610, "Python id": 611, "finished": true, "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877136.487, "dur": 76.777,
    "args": {
      "Python parent id": 611, "Python id": 612, "finished": true, "Ev Idx": 3512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877136.963, "dur": 0.335,
    "args": {
      "Python parent id": 612, "Python id": 613, "finished": true, "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877137.197, "dur": 0.045,
    "args": {
      "Python parent id": 613, "Python id": 614, "finished": true, "Ev Idx": 3514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877137.609, "dur": 0.303,
    "args": {
      "Python parent id": 612, "Python id": 615, "finished": true, "Ev Idx": 3515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877137.814, "dur": 0.042,
    "args": {
      "Python parent id": 615, "Python id": 616, "finished": true, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877138.206, "dur": 74.979,
    "args": {
      "Python parent id": 612, "Python id": 617, "finished": true, "Ev Idx": 3517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877138.828, "dur": 73.201,
    "args": {
      "Python parent id": 617, "Python id": 618, "finished": true, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877145.460, "dur": 16.103,
    "args": {
      "Python parent id": 618, "Python id": 619, "finished": true, "Ev Idx": 3519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877145.582, "dur": 0.094,
    "args": {
      "Python parent id": 619, "Python id": 620, "finished": true, "Ev Idx": 3520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877145.899, "dur": 15.539,
    "args": {
      "Python parent id": 619, "Python id": 621, "finished": true, "Ev Idx": 3521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877146.002, "dur": 6.638,
    "args": {
      "Python parent id": 621, "Python id": 622, "finished": true, "Ev Idx": 3522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877170.313, "dur": 8.254,
    "args": {
      "Python parent id": 618, "Python id": 623, "finished": true, "Ev Idx": 3523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877188.117, "dur": 8.220,
    "args": {
      "Python parent id": 618, "Python id": 624, "finished": true, "Ev Idx": 3524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877215.335, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 625, "finished": true, "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877215.744, "dur": 0.170,
    "args": {
      "Python parent id": 12, "Python id": 626, "finished": true, "Ev Idx": 3526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877216.181, "dur": 0.041,
    "args": {
      "Python parent id": 12, "Python id": 627, "finished": true, "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877216.415, "dur": 8.632,
    "args": {
      "Python parent id": 12, "Python id": 628, "finished": true, "Ev Idx": 3528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877226.057, "dur": 88.771,
    "args": {
      "Python parent id": 12, "Python id": 629, "finished": true, "Ev Idx": 3529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877226.371, "dur": 0.390,
    "args": {
      "Python parent id": 629, "Python id": 630, "finished": true, "Ev Idx": 3530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877226.629, "dur": 0.074,
    "args": {
      "Python parent id": 630, "Python id": 631, "finished": true, "Ev Idx": 3531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877227.038, "dur": 0.315,
    "args": {
      "Python parent id": 629, "Python id": 632, "finished": true, "Ev Idx": 3532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877227.634, "dur": 0.202,
    "args": {
      "Python parent id": 629, "Python id": 633, "finished": true, "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877228.058, "dur": 1.208,
    "args": {
      "Python parent id": 629, "Python id": 634, "finished": true, "Ev Idx": 3534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877229.804, "dur": 84.947,
    "args": {
      "Python parent id": 629, "Python id": 635, "finished": true, "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877230.498, "dur": 84.053,
    "args": {
      "Python parent id": 635, "Python id": 636, "finished": true, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877230.711, "dur": 83.761,
    "args": {
      "Python parent id": 636, "Python id": 637, "finished": true, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877234.867, "dur": 78.598,
    "args": {
      "Python parent id": 637, "Python id": 638, "finished": true, "Ev Idx": 3538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877235.131, "dur": 78.218,
    "args": {
      "Python parent id": 638, "Python id": 639, "finished": true, "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877235.685, "dur": 77.596,
    "args": {
      "Python parent id": 639, "Python id": 640, "finished": true, "Ev Idx": 3540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877236.129, "dur": 0.332,
    "args": {
      "Python parent id": 640, "Python id": 641, "finished": true, "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877236.363, "dur": 0.042,
    "args": {
      "Python parent id": 641, "Python id": 642, "finished": true, "Ev Idx": 3542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877236.752, "dur": 0.319,
    "args": {
      "Python parent id": 640, "Python id": 643, "finished": true, "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877236.973, "dur": 0.044,
    "args": {
      "Python parent id": 643, "Python id": 644, "finished": true, "Ev Idx": 3544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877237.361, "dur": 75.803,
    "args": {
      "Python parent id": 640, "Python id": 645, "finished": true, "Ev Idx": 3545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877237.997, "dur": 74.014,
    "args": {
      "Python parent id": 645, "Python id": 646, "finished": true, "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877244.590, "dur": 19.319,
    "args": {
      "Python parent id": 646, "Python id": 647, "finished": true, "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877244.716, "dur": 0.086,
    "args": {
      "Python parent id": 647, "Python id": 648, "finished": true, "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877245.081, "dur": 18.731,
    "args": {
      "Python parent id": 647, "Python id": 649, "finished": true, "Ev Idx": 3549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877245.183, "dur": 8.264,
    "args": {
      "Python parent id": 649, "Python id": 650, "finished": true, "Ev Idx": 3550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877271.514, "dur": 6.864,
    "args": {
      "Python parent id": 646, "Python id": 651, "finished": true, "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877287.710, "dur": 7.697,
    "args": {
      "Python parent id": 646, "Python id": 652, "finished": true, "Ev Idx": 3552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877315.444, "dur": 0.101,
    "args": {
      "Python parent id": 12, "Python id": 653, "finished": true, "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877315.860, "dur": 0.154,
    "args": {
      "Python parent id": 12, "Python id": 654, "finished": true, "Ev Idx": 3554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877316.258, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 655, "finished": true, "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877316.463, "dur": 9.775,
    "args": {
      "Python parent id": 12, "Python id": 656, "finished": true, "Ev Idx": 3556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877327.288, "dur": 88.697,
    "args": {
      "Python parent id": 12, "Python id": 657, "finished": true, "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877327.613, "dur": 0.421,
    "args": {
      "Python parent id": 657, "Python id": 658, "finished": true, "Ev Idx": 3558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877327.891, "dur": 0.082,
    "args": {
      "Python parent id": 658, "Python id": 659, "finished": true, "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877328.344, "dur": 0.296,
    "args": {
      "Python parent id": 657, "Python id": 660, "finished": true, "Ev Idx": 3560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877328.855, "dur": 0.237,
    "args": {
      "Python parent id": 657, "Python id": 661, "finished": true, "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877329.311, "dur": 1.249,
    "args": {
      "Python parent id": 657, "Python id": 662, "finished": true, "Ev Idx": 3562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877331.075, "dur": 84.834,
    "args": {
      "Python parent id": 657, "Python id": 663, "finished": true, "Ev Idx": 3563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877331.716, "dur": 84.029,
    "args": {
      "Python parent id": 663, "Python id": 664, "finished": true, "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877331.943, "dur": 83.745,
    "args": {
      "Python parent id": 664, "Python id": 665, "finished": true, "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877336.151, "dur": 78.555,
    "args": {
      "Python parent id": 665, "Python id": 666, "finished": true, "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877336.416, "dur": 78.178,
    "args": {
      "Python parent id": 666, "Python id": 667, "finished": true, "Ev Idx": 3567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877336.908, "dur": 77.610,
    "args": {
      "Python parent id": 667, "Python id": 668, "finished": true, "Ev Idx": 3568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877337.344, "dur": 0.366,
    "args": {
      "Python parent id": 668, "Python id": 669, "finished": true, "Ev Idx": 3569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877337.612, "dur": 0.044,
    "args": {
      "Python parent id": 669, "Python id": 670, "finished": true, "Ev Idx": 3570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877337.978, "dur": 0.313,
    "args": {
      "Python parent id": 668, "Python id": 671, "finished": true, "Ev Idx": 3571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877338.197, "dur": 0.040,
    "args": {
      "Python parent id": 671, "Python id": 672, "finished": true, "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877338.627, "dur": 75.812,
    "args": {
      "Python parent id": 668, "Python id": 673, "finished": true, "Ev Idx": 3573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877339.194, "dur": 74.103,
    "args": {
      "Python parent id": 673, "Python id": 674, "finished": true, "Ev Idx": 3574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877345.659, "dur": 16.452,
    "args": {
      "Python parent id": 674, "Python id": 675, "finished": true, "Ev Idx": 3575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877345.767, "dur": 0.084,
    "args": {
      "Python parent id": 675, "Python id": 676, "finished": true, "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877346.064, "dur": 15.951,
    "args": {
      "Python parent id": 675, "Python id": 677, "finished": true, "Ev Idx": 3577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877346.199, "dur": 7.555,
    "args": {
      "Python parent id": 677, "Python id": 678, "finished": true, "Ev Idx": 3578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877370.886, "dur": 7.829,
    "args": {
      "Python parent id": 674, "Python id": 679, "finished": true, "Ev Idx": 3579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877388.228, "dur": 8.981,
    "args": {
      "Python parent id": 674, "Python id": 680, "finished": true, "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877416.589, "dur": 0.107,
    "args": {
      "Python parent id": 12, "Python id": 681, "finished": true, "Ev Idx": 3581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877417.033, "dur": 0.148,
    "args": {
      "Python parent id": 12, "Python id": 682, "finished": true, "Ev Idx": 3582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877417.390, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 683, "finished": true, "Ev Idx": 3583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877417.597, "dur": 9.387,
    "args": {
      "Python parent id": 12, "Python id": 684, "finished": true, "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877427.960, "dur": 86.785,
    "args": {
      "Python parent id": 12, "Python id": 685, "finished": true, "Ev Idx": 3585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877428.260, "dur": 0.427,
    "args": {
      "Python parent id": 685, "Python id": 686, "finished": true, "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877428.545, "dur": 0.082,
    "args": {
      "Python parent id": 686, "Python id": 687, "finished": true, "Ev Idx": 3587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877428.940, "dur": 0.304,
    "args": {
      "Python parent id": 685, "Python id": 688, "finished": true, "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877429.469, "dur": 0.212,
    "args": {
      "Python parent id": 685, "Python id": 689, "finished": true, "Ev Idx": 3589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877429.910, "dur": 1.253,
    "args": {
      "Python parent id": 685, "Python id": 690, "finished": true, "Ev Idx": 3590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877431.702, "dur": 82.916,
    "args": {
      "Python parent id": 685, "Python id": 691, "finished": true, "Ev Idx": 3591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877432.352, "dur": 82.078,
    "args": {
      "Python parent id": 691, "Python id": 692, "finished": true, "Ev Idx": 3592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877432.588, "dur": 81.790,
    "args": {
      "Python parent id": 692, "Python id": 693, "finished": true, "Ev Idx": 3593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877436.687, "dur": 76.724,
    "args": {
      "Python parent id": 693, "Python id": 694, "finished": true, "Ev Idx": 3594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877436.914, "dur": 76.440,
    "args": {
      "Python parent id": 694, "Python id": 695, "finished": true, "Ev Idx": 3595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877437.411, "dur": 75.754,
    "args": {
      "Python parent id": 695, "Python id": 696, "finished": true, "Ev Idx": 3596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877437.826, "dur": 0.342,
    "args": {
      "Python parent id": 696, "Python id": 697, "finished": true, "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877438.065, "dur": 0.048,
    "args": {
      "Python parent id": 697, "Python id": 698, "finished": true, "Ev Idx": 3598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877438.436, "dur": 0.305,
    "args": {
      "Python parent id": 696, "Python id": 699, "finished": true, "Ev Idx": 3599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877438.640, "dur": 0.046,
    "args": {
      "Python parent id": 699, "Python id": 700, "finished": true, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877439.045, "dur": 74.049,
    "args": {
      "Python parent id": 696, "Python id": 701, "finished": true, "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877439.649, "dur": 72.300,
    "args": {
      "Python parent id": 701, "Python id": 702, "finished": true, "Ev Idx": 3602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877447.018, "dur": 18.119,
    "args": {
      "Python parent id": 702, "Python id": 703, "finished": true, "Ev Idx": 3603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877447.113, "dur": 0.091,
    "args": {
      "Python parent id": 703, "Python id": 704, "finished": true, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877447.464, "dur": 17.582,
    "args": {
      "Python parent id": 703, "Python id": 705, "finished": true, "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877447.598, "dur": 8.346,
    "args": {
      "Python parent id": 705, "Python id": 706, "finished": true, "Ev Idx": 3606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877472.537, "dur": 7.074,
    "args": {
      "Python parent id": 702, "Python id": 707, "finished": true, "Ev Idx": 3607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877487.882, "dur": 8.172,
    "args": {
      "Python parent id": 702, "Python id": 708, "finished": true, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877515.414, "dur": 0.103,
    "args": {
      "Python parent id": 12, "Python id": 709, "finished": true, "Ev Idx": 3609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877515.888, "dur": 0.153,
    "args": {
      "Python parent id": 12, "Python id": 710, "finished": true, "Ev Idx": 3610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877516.265, "dur": 0.038,
    "args": {
      "Python parent id": 12, "Python id": 711, "finished": true, "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877516.467, "dur": 9.035,
    "args": {
      "Python parent id": 12, "Python id": 712, "finished": true, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877526.584, "dur": 86.747,
    "args": {
      "Python parent id": 12, "Python id": 713, "finished": true, "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877526.931, "dur": 0.463,
    "args": {
      "Python parent id": 713, "Python id": 714, "finished": true, "Ev Idx": 3614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877527.231, "dur": 0.096,
    "args": {
      "Python parent id": 714, "Python id": 715, "finished": true, "Ev Idx": 3615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877527.642, "dur": 0.381,
    "args": {
      "Python parent id": 713, "Python id": 716, "finished": true, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877528.219, "dur": 0.230,
    "args": {
      "Python parent id": 713, "Python id": 717, "finished": true, "Ev Idx": 3617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877528.668, "dur": 1.205,
    "args": {
      "Python parent id": 713, "Python id": 718, "finished": true, "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877530.431, "dur": 82.772,
    "args": {
      "Python parent id": 713, "Python id": 719, "finished": true, "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877531.099, "dur": 81.964,
    "args": {
      "Python parent id": 719, "Python id": 720, "finished": true, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877531.308, "dur": 81.690,
    "args": {
      "Python parent id": 720, "Python id": 721, "finished": true, "Ev Idx": 3621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877535.516, "dur": 76.506,
    "args": {
      "Python parent id": 721, "Python id": 722, "finished": true, "Ev Idx": 3622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877535.807, "dur": 76.105,
    "args": {
      "Python parent id": 722, "Python id": 723, "finished": true, "Ev Idx": 3623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877536.309, "dur": 75.537,
    "args": {
      "Python parent id": 723, "Python id": 724, "finished": true, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877536.742, "dur": 0.367,
    "args": {
      "Python parent id": 724, "Python id": 725, "finished": true, "Ev Idx": 3625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877537.012, "dur": 0.041,
    "args": {
      "Python parent id": 725, "Python id": 726, "finished": true, "Ev Idx": 3626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877537.388, "dur": 0.346,
    "args": {
      "Python parent id": 724, "Python id": 727, "finished": true, "Ev Idx": 3627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877537.631, "dur": 0.054,
    "args": {
      "Python parent id": 727, "Python id": 728, "finished": true, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877538.082, "dur": 73.648,
    "args": {
      "Python parent id": 724, "Python id": 729, "finished": true, "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877538.693, "dur": 71.846,
    "args": {
      "Python parent id": 729, "Python id": 730, "finished": true, "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877545.270, "dur": 17.683,
    "args": {
      "Python parent id": 730, "Python id": 731, "finished": true, "Ev Idx": 3631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877545.393, "dur": 0.089,
    "args": {
      "Python parent id": 731, "Python id": 732, "finished": true, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877545.721, "dur": 17.153,
    "args": {
      "Python parent id": 731, "Python id": 733, "finished": true, "Ev Idx": 3633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877545.899, "dur": 8.485,
    "args": {
      "Python parent id": 733, "Python id": 734, "finished": true, "Ev Idx": 3634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877570.392, "dur": 7.673,
    "args": {
      "Python parent id": 730, "Python id": 735, "finished": true, "Ev Idx": 3635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877586.694, "dur": 7.899,
    "args": {
      "Python parent id": 730, "Python id": 736, "finished": true, "Ev Idx": 3636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877613.985, "dur": 0.121,
    "args": {
      "Python parent id": 12, "Python id": 737, "finished": true, "Ev Idx": 3637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877614.447, "dur": 0.149,
    "args": {
      "Python parent id": 12, "Python id": 738, "finished": true, "Ev Idx": 3638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877614.828, "dur": 0.049,
    "args": {
      "Python parent id": 12, "Python id": 739, "finished": true, "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877615.059, "dur": 9.003,
    "args": {
      "Python parent id": 12, "Python id": 740, "finished": true, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877625.084, "dur": 85.524,
    "args": {
      "Python parent id": 12, "Python id": 741, "finished": true, "Ev Idx": 3641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877625.382, "dur": 0.430,
    "args": {
      "Python parent id": 741, "Python id": 742, "finished": true, "Ev Idx": 3642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877625.673, "dur": 0.078,
    "args": {
      "Python parent id": 742, "Python id": 743, "finished": true, "Ev Idx": 3643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877626.057, "dur": 0.376,
    "args": {
      "Python parent id": 741, "Python id": 744, "finished": true, "Ev Idx": 3644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877626.694, "dur": 0.248,
    "args": {
      "Python parent id": 741, "Python id": 745, "finished": true, "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877627.155, "dur": 1.229,
    "args": {
      "Python parent id": 741, "Python id": 746, "finished": true, "Ev Idx": 3646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877628.923, "dur": 81.576,
    "args": {
      "Python parent id": 741, "Python id": 747, "finished": true, "Ev Idx": 3647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877629.662, "dur": 80.673,
    "args": {
      "Python parent id": 747, "Python id": 748, "finished": true, "Ev Idx": 3648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877629.866, "dur": 80.389,
    "args": {
      "Python parent id": 748, "Python id": 749, "finished": true, "Ev Idx": 3649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877634.346, "dur": 74.862,
    "args": {
      "Python parent id": 749, "Python id": 750, "finished": true, "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877634.607, "dur": 74.534,
    "args": {
      "Python parent id": 750, "Python id": 751, "finished": true, "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877635.113, "dur": 73.892,
    "args": {
      "Python parent id": 751, "Python id": 752, "finished": true, "Ev Idx": 3652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877635.560, "dur": 0.346,
    "args": {
      "Python parent id": 752, "Python id": 753, "finished": true, "Ev Idx": 3653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877635.795, "dur": 0.063,
    "args": {
      "Python parent id": 753, "Python id": 754, "finished": true, "Ev Idx": 3654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877636.194, "dur": 0.295,
    "args": {
      "Python parent id": 752, "Python id": 755, "finished": true, "Ev Idx": 3655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877636.395, "dur": 0.041,
    "args": {
      "Python parent id": 755, "Python id": 756, "finished": true, "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877636.776, "dur": 72.109,
    "args": {
      "Python parent id": 752, "Python id": 757, "finished": true, "Ev Idx": 3657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877637.403, "dur": 70.348,
    "args": {
      "Python parent id": 757, "Python id": 758, "finished": true, "Ev Idx": 3658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877643.962, "dur": 17.279,
    "args": {
      "Python parent id": 758, "Python id": 759, "finished": true, "Ev Idx": 3659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877644.056, "dur": 0.119,
    "args": {
      "Python parent id": 759, "Python id": 760, "finished": true, "Ev Idx": 3660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877644.416, "dur": 16.729,
    "args": {
      "Python parent id": 759, "Python id": 761, "finished": true, "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877644.526, "dur": 7.357,
    "args": {
      "Python parent id": 761, "Python id": 762, "finished": true, "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877669.019, "dur": 7.133,
    "args": {
      "Python parent id": 758, "Python id": 763, "finished": true, "Ev Idx": 3663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877684.467, "dur": 8.346,
    "args": {
      "Python parent id": 758, "Python id": 764, "finished": true, "Ev Idx": 3664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877711.284, "dur": 0.109,
    "args": {
      "Python parent id": 12, "Python id": 765, "finished": true, "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877711.736, "dur": 0.172,
    "args": {
      "Python parent id": 12, "Python id": 766, "finished": true, "Ev Idx": 3666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877712.157, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 767, "finished": true, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877712.395, "dur": 8.626,
    "args": {
      "Python parent id": 12, "Python id": 768, "finished": true, "Ev Idx": 3668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877722.329, "dur": 97.954,
    "args": {
      "Python parent id": 12, "Python id": 769, "finished": true, "Ev Idx": 3669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877722.717, "dur": 0.432,
    "args": {
      "Python parent id": 769, "Python id": 770, "finished": true, "Ev Idx": 3670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877723.009, "dur": 0.075,
    "args": {
      "Python parent id": 770, "Python id": 771, "finished": true, "Ev Idx": 3671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877723.581, "dur": 0.318,
    "args": {
      "Python parent id": 769, "Python id": 772, "finished": true, "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877724.185, "dur": 0.225,
    "args": {
      "Python parent id": 769, "Python id": 773, "finished": true, "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877724.747, "dur": 1.262,
    "args": {
      "Python parent id": 769, "Python id": 774, "finished": true, "Ev Idx": 3674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877726.524, "dur": 93.627,
    "args": {
      "Python parent id": 769, "Python id": 775, "finished": true, "Ev Idx": 3675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877727.357, "dur": 92.613,
    "args": {
      "Python parent id": 775, "Python id": 776, "finished": true, "Ev Idx": 3676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877727.557, "dur": 92.354,
    "args": {
      "Python parent id": 776, "Python id": 777, "finished": true, "Ev Idx": 3677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877736.501, "dur": 82.359,
    "args": {
      "Python parent id": 777, "Python id": 778, "finished": true, "Ev Idx": 3678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877736.798, "dur": 81.982,
    "args": {
      "Python parent id": 778, "Python id": 779, "finished": true, "Ev Idx": 3679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877737.334, "dur": 81.361,
    "args": {
      "Python parent id": 779, "Python id": 780, "finished": true, "Ev Idx": 3680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877737.892, "dur": 0.377,
    "args": {
      "Python parent id": 780, "Python id": 781, "finished": true, "Ev Idx": 3681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877738.152, "dur": 0.069,
    "args": {
      "Python parent id": 781, "Python id": 782, "finished": true, "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877738.557, "dur": 0.312,
    "args": {
      "Python parent id": 780, "Python id": 783, "finished": true, "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877738.771, "dur": 0.040,
    "args": {
      "Python parent id": 783, "Python id": 784, "finished": true, "Ev Idx": 3684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877739.163, "dur": 79.429,
    "args": {
      "Python parent id": 780, "Python id": 785, "finished": true, "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877739.954, "dur": 77.293,
    "args": {
      "Python parent id": 785, "Python id": 786, "finished": true, "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877747.412, "dur": 16.781,
    "args": {
      "Python parent id": 786, "Python id": 787, "finished": true, "Ev Idx": 3687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877747.512, "dur": 0.086,
    "args": {
      "Python parent id": 787, "Python id": 788, "finished": true, "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877747.845, "dur": 16.247,
    "args": {
      "Python parent id": 787, "Python id": 789, "finished": true, "Ev Idx": 3689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877747.977, "dur": 6.975,
    "args": {
      "Python parent id": 789, "Python id": 790, "finished": true, "Ev Idx": 3690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877772.794, "dur": 9.305,
    "args": {
      "Python parent id": 786, "Python id": 791, "finished": true, "Ev Idx": 3691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877792.431, "dur": 7.913,
    "args": {
      "Python parent id": 786, "Python id": 792, "finished": true, "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877820.992, "dur": 0.130,
    "args": {
      "Python parent id": 12, "Python id": 793, "finished": true, "Ev Idx": 3693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877821.466, "dur": 0.155,
    "args": {
      "Python parent id": 12, "Python id": 794, "finished": true, "Ev Idx": 3694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877821.897, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 795, "finished": true, "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877822.103, "dur": 9.566,
    "args": {
      "Python parent id": 12, "Python id": 796, "finished": true, "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877832.843, "dur": 92.668,
    "args": {
      "Python parent id": 12, "Python id": 797, "finished": true, "Ev Idx": 3697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877833.183, "dur": 0.478,
    "args": {
      "Python parent id": 797, "Python id": 798, "finished": true, "Ev Idx": 3698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877833.486, "dur": 0.102,
    "args": {
      "Python parent id": 798, "Python id": 799, "finished": true, "Ev Idx": 3699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877833.969, "dur": 0.315,
    "args": {
      "Python parent id": 797, "Python id": 800, "finished": true, "Ev Idx": 3700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877834.537, "dur": 0.225,
    "args": {
      "Python parent id": 797, "Python id": 801, "finished": true, "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877835.016, "dur": 1.246,
    "args": {
      "Python parent id": 797, "Python id": 802, "finished": true, "Ev Idx": 3702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877836.831, "dur": 88.581,
    "args": {
      "Python parent id": 797, "Python id": 803, "finished": true, "Ev Idx": 3703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877837.532, "dur": 87.694,
    "args": {
      "Python parent id": 803, "Python id": 804, "finished": true, "Ev Idx": 3704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877837.801, "dur": 87.344,
    "args": {
      "Python parent id": 804, "Python id": 805, "finished": true, "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877842.315, "dur": 81.775,
    "args": {
      "Python parent id": 805, "Python id": 806, "finished": true, "Ev Idx": 3706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877842.648, "dur": 81.374,
    "args": {
      "Python parent id": 806, "Python id": 807, "finished": true, "Ev Idx": 3707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877843.134, "dur": 80.707,
    "args": {
      "Python parent id": 807, "Python id": 808, "finished": true, "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877843.604, "dur": 0.319,
    "args": {
      "Python parent id": 808, "Python id": 809, "finished": true, "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877843.830, "dur": 0.039,
    "args": {
      "Python parent id": 809, "Python id": 810, "finished": true, "Ev Idx": 3710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877844.232, "dur": 0.356,
    "args": {
      "Python parent id": 808, "Python id": 811, "finished": true, "Ev Idx": 3711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877844.483, "dur": 0.048,
    "args": {
      "Python parent id": 811, "Python id": 812, "finished": true, "Ev Idx": 3712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877844.888, "dur": 78.818,
    "args": {
      "Python parent id": 808, "Python id": 813, "finished": true, "Ev Idx": 3713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877845.491, "dur": 77.027,
    "args": {
      "Python parent id": 813, "Python id": 814, "finished": true, "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877852.224, "dur": 18.503,
    "args": {
      "Python parent id": 814, "Python id": 815, "finished": true, "Ev Idx": 3715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877852.328, "dur": 0.093,
    "args": {
      "Python parent id": 815, "Python id": 816, "finished": true, "Ev Idx": 3716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877852.718, "dur": 17.866,
    "args": {
      "Python parent id": 815, "Python id": 817, "finished": true, "Ev Idx": 3717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877852.823, "dur": 6.864,
    "args": {
      "Python parent id": 817, "Python id": 818, "finished": true, "Ev Idx": 3718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877878.773, "dur": 6.923,
    "args": {
      "Python parent id": 814, "Python id": 819, "finished": true, "Ev Idx": 3719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740877896.997, "dur": 8.413,
    "args": {
      "Python parent id": 814, "Python id": 820, "finished": true, "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740877926.152, "dur": 0.096,
    "args": {
      "Python parent id": 12, "Python id": 821, "finished": true, "Ev Idx": 3721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877926.594, "dur": 0.117,
    "args": {
      "Python parent id": 12, "Python id": 822, "finished": true, "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740877926.987, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 823, "finished": true, "Ev Idx": 3723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740877927.192, "dur": 9.121,
    "args": {
      "Python parent id": 12, "Python id": 824, "finished": true, "Ev Idx": 3724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740877937.331, "dur": 115.618,
    "args": {
      "Python parent id": 12, "Python id": 825, "finished": true, "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877937.676, "dur": 0.498,
    "args": {
      "Python parent id": 825, "Python id": 826, "finished": true, "Ev Idx": 3726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877938.016, "dur": 0.094,
    "args": {
      "Python parent id": 826, "Python id": 827, "finished": true, "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877938.447, "dur": 0.420,
    "args": {
      "Python parent id": 825, "Python id": 828, "finished": true, "Ev Idx": 3728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740877939.090, "dur": 0.227,
    "args": {
      "Python parent id": 825, "Python id": 829, "finished": true, "Ev Idx": 3729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740877939.525, "dur": 1.245,
    "args": {
      "Python parent id": 825, "Python id": 830, "finished": true, "Ev Idx": 3730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740877941.321, "dur": 111.513,
    "args": {
      "Python parent id": 825, "Python id": 831, "finished": true, "Ev Idx": 3731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740877942.039, "dur": 110.652,
    "args": {
      "Python parent id": 831, "Python id": 832, "finished": true, "Ev Idx": 3732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877942.263, "dur": 110.350,
    "args": {
      "Python parent id": 832, "Python id": 833, "finished": true, "Ev Idx": 3733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740877946.454, "dur": 105.159,
    "args": {
      "Python parent id": 833, "Python id": 834, "finished": true, "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740877946.712, "dur": 104.834,
    "args": {
      "Python parent id": 834, "Python id": 835, "finished": true, "Ev Idx": 3735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740877947.251, "dur": 104.166,
    "args": {
      "Python parent id": 835, "Python id": 836, "finished": true, "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877947.760, "dur": 0.338,
    "args": {
      "Python parent id": 836, "Python id": 837, "finished": true, "Ev Idx": 3737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877947.990, "dur": 0.050,
    "args": {
      "Python parent id": 837, "Python id": 838, "finished": true, "Ev Idx": 3738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740877948.370, "dur": 0.333,
    "args": {
      "Python parent id": 836, "Python id": 839, "finished": true, "Ev Idx": 3739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740877948.598, "dur": 0.046,
    "args": {
      "Python parent id": 839, "Python id": 840, "finished": true, "Ev Idx": 3740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877949.035, "dur": 102.291,
    "args": {
      "Python parent id": 836, "Python id": 841, "finished": true, "Ev Idx": 3741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740877949.633, "dur": 100.502,
    "args": {
      "Python parent id": 841, "Python id": 842, "finished": true, "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740877983.720, "dur": 17.687,
    "args": {
      "Python parent id": 842, "Python id": 843, "finished": true, "Ev Idx": 3743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740877983.865, "dur": 0.091,
    "args": {
      "Python parent id": 843, "Python id": 844, "finished": true, "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740877984.232, "dur": 17.043,
    "args": {
      "Python parent id": 843, "Python id": 845, "finished": true, "Ev Idx": 3745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740877984.340, "dur": 8.404,
    "args": {
      "Python parent id": 845, "Python id": 846, "finished": true, "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878009.070, "dur": 7.501,
    "args": {
      "Python parent id": 842, "Python id": 847, "finished": true, "Ev Idx": 3747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878025.675, "dur": 8.833,
    "args": {
      "Python parent id": 842, "Python id": 848, "finished": true, "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878053.668, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 849, "finished": true, "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878054.084, "dur": 0.141,
    "args": {
      "Python parent id": 12, "Python id": 850, "finished": true, "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878054.449, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 851, "finished": true, "Ev Idx": 3751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878054.656, "dur": 9.769,
    "args": {
      "Python parent id": 12, "Python id": 852, "finished": true, "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878065.375, "dur": 145.257,
    "args": {
      "Python parent id": 12, "Python id": 853, "finished": true, "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878065.729, "dur": 0.492,
    "args": {
      "Python parent id": 853, "Python id": 854, "finished": true, "Ev Idx": 3754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878066.070, "dur": 0.076,
    "args": {
      "Python parent id": 854, "Python id": 855, "finished": true, "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878066.492, "dur": 0.317,
    "args": {
      "Python parent id": 853, "Python id": 856, "finished": true, "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878067.047, "dur": 0.209,
    "args": {
      "Python parent id": 853, "Python id": 857, "finished": true, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878067.482, "dur": 1.167,
    "args": {
      "Python parent id": 853, "Python id": 858, "finished": true, "Ev Idx": 3758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878102.307, "dur": 108.200,
    "args": {
      "Python parent id": 853, "Python id": 859, "finished": true, "Ev Idx": 3759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878103.996, "dur": 106.302,
    "args": {
      "Python parent id": 859, "Python id": 860, "finished": true, "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878104.487, "dur": 105.736,
    "args": {
      "Python parent id": 860, "Python id": 861, "finished": true, "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878112.785, "dur": 95.840,
    "args": {
      "Python parent id": 861, "Python id": 862, "finished": true, "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878113.102, "dur": 95.457,
    "args": {
      "Python parent id": 862, "Python id": 863, "finished": true, "Ev Idx": 3763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878114.026, "dur": 94.319,
    "args": {
      "Python parent id": 863, "Python id": 864, "finished": true, "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878114.640, "dur": 0.423,
    "args": {
      "Python parent id": 864, "Python id": 865, "finished": true, "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878114.928, "dur": 0.044,
    "args": {
      "Python parent id": 865, "Python id": 866, "finished": true, "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878115.429, "dur": 0.341,
    "args": {
      "Python parent id": 864, "Python id": 867, "finished": true, "Ev Idx": 3767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878115.670, "dur": 0.044,
    "args": {
      "Python parent id": 867, "Python id": 868, "finished": true, "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878116.157, "dur": 92.073,
    "args": {
      "Python parent id": 864, "Python id": 869, "finished": true, "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878116.950, "dur": 89.856,
    "args": {
      "Python parent id": 869, "Python id": 870, "finished": true, "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878126.614, "dur": 25.845,
    "args": {
      "Python parent id": 870, "Python id": 871, "finished": true, "Ev Idx": 3771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878126.747, "dur": 0.162,
    "args": {
      "Python parent id": 871, "Python id": 872, "finished": true, "Ev Idx": 3772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878127.310, "dur": 24.952,
    "args": {
      "Python parent id": 871, "Python id": 873, "finished": true, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878127.494, "dur": 10.371,
    "args": {
      "Python parent id": 873, "Python id": 874, "finished": true, "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878161.677, "dur": 7.976,
    "args": {
      "Python parent id": 870, "Python id": 875, "finished": true, "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878179.397, "dur": 10.149,
    "args": {
      "Python parent id": 870, "Python id": 876, "finished": true, "Ev Idx": 3776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878211.593, "dur": 0.122,
    "args": {
      "Python parent id": 12, "Python id": 877, "finished": true, "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878212.262, "dur": 0.277,
    "args": {
      "Python parent id": 12, "Python id": 878, "finished": true, "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878212.870, "dur": 0.049,
    "args": {
      "Python parent id": 12, "Python id": 879, "finished": true, "Ev Idx": 3779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878213.110, "dur": 11.594,
    "args": {
      "Python parent id": 12, "Python id": 880, "finished": true, "Ev Idx": 3780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878225.960, "dur": 91.325,
    "args": {
      "Python parent id": 12, "Python id": 881, "finished": true, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878226.331, "dur": 0.460,
    "args": {
      "Python parent id": 881, "Python id": 882, "finished": true, "Ev Idx": 3782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878226.641, "dur": 0.104,
    "args": {
      "Python parent id": 882, "Python id": 883, "finished": true, "Ev Idx": 3783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878227.068, "dur": 0.464,
    "args": {
      "Python parent id": 881, "Python id": 884, "finished": true, "Ev Idx": 3784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878227.864, "dur": 0.290,
    "args": {
      "Python parent id": 881, "Python id": 885, "finished": true, "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878228.445, "dur": 1.317,
    "args": {
      "Python parent id": 881, "Python id": 886, "finished": true, "Ev Idx": 3786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878230.488, "dur": 86.692,
    "args": {
      "Python parent id": 881, "Python id": 887, "finished": true, "Ev Idx": 3787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878231.437, "dur": 85.536,
    "args": {
      "Python parent id": 887, "Python id": 888, "finished": true, "Ev Idx": 3788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878231.771, "dur": 85.147,
    "args": {
      "Python parent id": 888, "Python id": 889, "finished": true, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878237.139, "dur": 78.636,
    "args": {
      "Python parent id": 889, "Python id": 890, "finished": true, "Ev Idx": 3790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878237.483, "dur": 78.228,
    "args": {
      "Python parent id": 890, "Python id": 891, "finished": true, "Ev Idx": 3791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878238.124, "dur": 77.426,
    "args": {
      "Python parent id": 891, "Python id": 892, "finished": true, "Ev Idx": 3792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878238.717, "dur": 0.365,
    "args": {
      "Python parent id": 892, "Python id": 893, "finished": true, "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878238.976, "dur": 0.047,
    "args": {
      "Python parent id": 893, "Python id": 894, "finished": true, "Ev Idx": 3794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878239.394, "dur": 0.321,
    "args": {
      "Python parent id": 892, "Python id": 895, "finished": true, "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878239.616, "dur": 0.042,
    "args": {
      "Python parent id": 895, "Python id": 896, "finished": true, "Ev Idx": 3796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878240.075, "dur": 75.321,
    "args": {
      "Python parent id": 892, "Python id": 897, "finished": true, "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878240.656, "dur": 73.451,
    "args": {
      "Python parent id": 897, "Python id": 898, "finished": true, "Ev Idx": 3798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878247.684, "dur": 16.810,
    "args": {
      "Python parent id": 898, "Python id": 899, "finished": true, "Ev Idx": 3799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878247.833, "dur": 0.142,
    "args": {
      "Python parent id": 899, "Python id": 900, "finished": true, "Ev Idx": 3800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878248.380, "dur": 16.020,
    "args": {
      "Python parent id": 899, "Python id": 901, "finished": true, "Ev Idx": 3801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878248.503, "dur": 7.360,
    "args": {
      "Python parent id": 901, "Python id": 902, "finished": true, "Ev Idx": 3802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878272.786, "dur": 7.333,
    "args": {
      "Python parent id": 898, "Python id": 903, "finished": true, "Ev Idx": 3803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878289.022, "dur": 8.168,
    "args": {
      "Python parent id": 898, "Python id": 904, "finished": true, "Ev Idx": 3804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878318.015, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 905, "finished": true, "Ev Idx": 3805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878318.488, "dur": 0.191,
    "args": {
      "Python parent id": 12, "Python id": 906, "finished": true, "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878318.949, "dur": 0.043,
    "args": {
      "Python parent id": 12, "Python id": 907, "finished": true, "Ev Idx": 3807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878319.186, "dur": 8.805,
    "args": {
      "Python parent id": 12, "Python id": 908, "finished": true, "Ev Idx": 3808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878329.063, "dur": 88.086,
    "args": {
      "Python parent id": 12, "Python id": 909, "finished": true, "Ev Idx": 3809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878329.457, "dur": 0.498,
    "args": {
      "Python parent id": 909, "Python id": 910, "finished": true, "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878329.786, "dur": 0.126,
    "args": {
      "Python parent id": 910, "Python id": 911, "finished": true, "Ev Idx": 3811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878330.291, "dur": 0.369,
    "args": {
      "Python parent id": 909, "Python id": 912, "finished": true, "Ev Idx": 3812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878330.924, "dur": 0.280,
    "args": {
      "Python parent id": 909, "Python id": 913, "finished": true, "Ev Idx": 3813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878331.516, "dur": 1.356,
    "args": {
      "Python parent id": 909, "Python id": 914, "finished": true, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878333.484, "dur": 83.560,
    "args": {
      "Python parent id": 909, "Python id": 915, "finished": true, "Ev Idx": 3815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878334.221, "dur": 82.676,
    "args": {
      "Python parent id": 915, "Python id": 916, "finished": true, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878334.550, "dur": 82.296,
    "args": {
      "Python parent id": 916, "Python id": 917, "finished": true, "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878339.104, "dur": 76.662,
    "args": {
      "Python parent id": 917, "Python id": 918, "finished": true, "Ev Idx": 3818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878339.453, "dur": 76.216,
    "args": {
      "Python parent id": 918, "Python id": 919, "finished": true, "Ev Idx": 3819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878340.081, "dur": 75.469,
    "args": {
      "Python parent id": 919, "Python id": 920, "finished": true, "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878340.699, "dur": 0.430,
    "args": {
      "Python parent id": 920, "Python id": 921, "finished": true, "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878341.033, "dur": 0.042,
    "args": {
      "Python parent id": 921, "Python id": 922, "finished": true, "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878341.491, "dur": 0.347,
    "args": {
      "Python parent id": 920, "Python id": 923, "finished": true, "Ev Idx": 3823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878341.740, "dur": 0.040,
    "args": {
      "Python parent id": 923, "Python id": 924, "finished": true, "Ev Idx": 3824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878342.228, "dur": 73.226,
    "args": {
      "Python parent id": 920, "Python id": 925, "finished": true, "Ev Idx": 3825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878342.885, "dur": 71.277,
    "args": {
      "Python parent id": 925, "Python id": 926, "finished": true, "Ev Idx": 3826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878349.806, "dur": 17.298,
    "args": {
      "Python parent id": 926, "Python id": 927, "finished": true, "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878349.944, "dur": 0.127,
    "args": {
      "Python parent id": 927, "Python id": 928, "finished": true, "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878350.343, "dur": 16.701,
    "args": {
      "Python parent id": 927, "Python id": 929, "finished": true, "Ev Idx": 3829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878350.455, "dur": 7.562,
    "args": {
      "Python parent id": 929, "Python id": 930, "finished": true, "Ev Idx": 3830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878374.517, "dur": 6.384,
    "args": {
      "Python parent id": 926, "Python id": 931, "finished": true, "Ev Idx": 3831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878389.631, "dur": 8.294,
    "args": {
      "Python parent id": 926, "Python id": 932, "finished": true, "Ev Idx": 3832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878417.801, "dur": 0.111,
    "args": {
      "Python parent id": 12, "Python id": 933, "finished": true, "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878418.247, "dur": 0.236,
    "args": {
      "Python parent id": 12, "Python id": 934, "finished": true, "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878418.720, "dur": 0.041,
    "args": {
      "Python parent id": 12, "Python id": 935, "finished": true, "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878418.927, "dur": 9.791,
    "args": {
      "Python parent id": 12, "Python id": 936, "finished": true, "Ev Idx": 3836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878429.741, "dur": 86.547,
    "args": {
      "Python parent id": 12, "Python id": 937, "finished": true, "Ev Idx": 3837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878430.094, "dur": 0.507,
    "args": {
      "Python parent id": 937, "Python id": 938, "finished": true, "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878430.458, "dur": 0.077,
    "args": {
      "Python parent id": 938, "Python id": 939, "finished": true, "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878430.906, "dur": 0.378,
    "args": {
      "Python parent id": 937, "Python id": 940, "finished": true, "Ev Idx": 3840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878431.536, "dur": 0.257,
    "args": {
      "Python parent id": 937, "Python id": 941, "finished": true, "Ev Idx": 3841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878432.021, "dur": 1.253,
    "args": {
      "Python parent id": 937, "Python id": 942, "finished": true, "Ev Idx": 3842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878433.862, "dur": 82.330,
    "args": {
      "Python parent id": 937, "Python id": 943, "finished": true, "Ev Idx": 3843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878434.692, "dur": 81.332,
    "args": {
      "Python parent id": 943, "Python id": 944, "finished": true, "Ev Idx": 3844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878434.964, "dur": 80.977,
    "args": {
      "Python parent id": 944, "Python id": 945, "finished": true, "Ev Idx": 3845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878439.562, "dur": 75.343,
    "args": {
      "Python parent id": 945, "Python id": 946, "finished": true, "Ev Idx": 3846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878439.833, "dur": 74.983,
    "args": {
      "Python parent id": 946, "Python id": 947, "finished": true, "Ev Idx": 3847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878440.400, "dur": 74.266,
    "args": {
      "Python parent id": 947, "Python id": 948, "finished": true, "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878440.933, "dur": 0.394,
    "args": {
      "Python parent id": 948, "Python id": 949, "finished": true, "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878441.205, "dur": 0.075,
    "args": {
      "Python parent id": 949, "Python id": 950, "finished": true, "Ev Idx": 3850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878441.651, "dur": 0.375,
    "args": {
      "Python parent id": 948, "Python id": 951, "finished": true, "Ev Idx": 3851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878441.929, "dur": 0.042,
    "args": {
      "Python parent id": 951, "Python id": 952, "finished": true, "Ev Idx": 3852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878442.375, "dur": 72.200,
    "args": {
      "Python parent id": 948, "Python id": 953, "finished": true, "Ev Idx": 3853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878443.053, "dur": 70.266,
    "args": {
      "Python parent id": 953, "Python id": 954, "finished": true, "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878449.921, "dur": 16.529,
    "args": {
      "Python parent id": 954, "Python id": 955, "finished": true, "Ev Idx": 3855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878450.043, "dur": 0.094,
    "args": {
      "Python parent id": 955, "Python id": 956, "finished": true, "Ev Idx": 3856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878450.358, "dur": 16.021,
    "args": {
      "Python parent id": 955, "Python id": 957, "finished": true, "Ev Idx": 3857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878450.463, "dur": 7.043,
    "args": {
      "Python parent id": 957, "Python id": 958, "finished": true, "Ev Idx": 3858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878474.702, "dur": 7.057,
    "args": {
      "Python parent id": 954, "Python id": 959, "finished": true, "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878490.342, "dur": 7.986,
    "args": {
      "Python parent id": 954, "Python id": 960, "finished": true, "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878517.028, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 961, "finished": true, "Ev Idx": 3861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878517.487, "dur": 0.153,
    "args": {
      "Python parent id": 12, "Python id": 962, "finished": true, "Ev Idx": 3862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878517.948, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 963, "finished": true, "Ev Idx": 3863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878518.153, "dur": 8.811,
    "args": {
      "Python parent id": 12, "Python id": 964, "finished": true, "Ev Idx": 3864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878527.917, "dur": 88.124,
    "args": {
      "Python parent id": 12, "Python id": 965, "finished": true, "Ev Idx": 3865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878528.253, "dur": 0.412,
    "args": {
      "Python parent id": 965, "Python id": 966, "finished": true, "Ev Idx": 3866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878528.524, "dur": 0.076,
    "args": {
      "Python parent id": 966, "Python id": 967, "finished": true, "Ev Idx": 3867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878528.918, "dur": 0.391,
    "args": {
      "Python parent id": 965, "Python id": 968, "finished": true, "Ev Idx": 3868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878529.533, "dur": 0.245,
    "args": {
      "Python parent id": 965, "Python id": 969, "finished": true, "Ev Idx": 3869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878530.028, "dur": 1.225,
    "args": {
      "Python parent id": 965, "Python id": 970, "finished": true, "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878531.880, "dur": 84.067,
    "args": {
      "Python parent id": 965, "Python id": 971, "finished": true, "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878532.728, "dur": 83.023,
    "args": {
      "Python parent id": 971, "Python id": 972, "finished": true, "Ev Idx": 3872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878532.974, "dur": 82.717,
    "args": {
      "Python parent id": 972, "Python id": 973, "finished": true, "Ev Idx": 3873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878537.593, "dur": 76.958,
    "args": {
      "Python parent id": 973, "Python id": 974, "finished": true, "Ev Idx": 3874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878537.856, "dur": 76.584,
    "args": {
      "Python parent id": 974, "Python id": 975, "finished": true, "Ev Idx": 3875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878538.416, "dur": 75.943,
    "args": {
      "Python parent id": 975, "Python id": 976, "finished": true, "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878538.930, "dur": 0.395,
    "args": {
      "Python parent id": 976, "Python id": 977, "finished": true, "Ev Idx": 3877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878539.223, "dur": 0.042,
    "args": {
      "Python parent id": 977, "Python id": 978, "finished": true, "Ev Idx": 3878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878539.584, "dur": 0.390,
    "args": {
      "Python parent id": 976, "Python id": 979, "finished": true, "Ev Idx": 3879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878539.834, "dur": 0.095,
    "args": {
      "Python parent id": 979, "Python id": 980, "finished": true, "Ev Idx": 3880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878540.282, "dur": 73.980,
    "args": {
      "Python parent id": 976, "Python id": 981, "finished": true, "Ev Idx": 3881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878540.912, "dur": 72.064,
    "args": {
      "Python parent id": 981, "Python id": 982, "finished": true, "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878547.519, "dur": 17.356,
    "args": {
      "Python parent id": 982, "Python id": 983, "finished": true, "Ev Idx": 3883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878547.643, "dur": 0.092,
    "args": {
      "Python parent id": 983, "Python id": 984, "finished": true, "Ev Idx": 3884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878548.023, "dur": 16.741,
    "args": {
      "Python parent id": 983, "Python id": 985, "finished": true, "Ev Idx": 3885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878548.180, "dur": 7.764,
    "args": {
      "Python parent id": 985, "Python id": 986, "finished": true, "Ev Idx": 3886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878572.271, "dur": 7.157,
    "args": {
      "Python parent id": 982, "Python id": 987, "finished": true, "Ev Idx": 3887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878587.951, "dur": 7.943,
    "args": {
      "Python parent id": 982, "Python id": 988, "finished": true, "Ev Idx": 3888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878616.764, "dur": 0.103,
    "args": {
      "Python parent id": 12, "Python id": 989, "finished": true, "Ev Idx": 3889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878617.200, "dur": 0.141,
    "args": {
      "Python parent id": 12, "Python id": 990, "finished": true, "Ev Idx": 3890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878617.606, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 991, "finished": true, "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878617.823, "dur": 9.782,
    "args": {
      "Python parent id": 12, "Python id": 992, "finished": true, "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878628.657, "dur": 87.079,
    "args": {
      "Python parent id": 12, "Python id": 993, "finished": true, "Ev Idx": 3893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878628.971, "dur": 0.466,
    "args": {
      "Python parent id": 993, "Python id": 994, "finished": true, "Ev Idx": 3894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878629.270, "dur": 0.098,
    "args": {
      "Python parent id": 994, "Python id": 995, "finished": true, "Ev Idx": 3895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878629.757, "dur": 0.401,
    "args": {
      "Python parent id": 993, "Python id": 996, "finished": true, "Ev Idx": 3896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878630.430, "dur": 0.277,
    "args": {
      "Python parent id": 993, "Python id": 997, "finished": true, "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878630.969, "dur": 1.258,
    "args": {
      "Python parent id": 993, "Python id": 998, "finished": true, "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878632.819, "dur": 82.841,
    "args": {
      "Python parent id": 993, "Python id": 999, "finished": true, "Ev Idx": 3899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878633.618, "dur": 81.888,
    "args": {
      "Python parent id": 999, "Python id": 1000, "finished": true, "Ev Idx": 3900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878633.811, "dur": 81.614,
    "args": {
      "Python parent id": 1000, "Python id": 1001, "finished": true, "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878638.039, "dur": 76.330,
    "args": {
      "Python parent id": 1001, "Python id": 1002, "finished": true, "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878638.339, "dur": 75.907,
    "args": {
      "Python parent id": 1002, "Python id": 1003, "finished": true, "Ev Idx": 3903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878638.900, "dur": 75.265,
    "args": {
      "Python parent id": 1003, "Python id": 1004, "finished": true, "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878639.446, "dur": 0.363,
    "args": {
      "Python parent id": 1004, "Python id": 1005, "finished": true, "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878639.701, "dur": 0.049,
    "args": {
      "Python parent id": 1005, "Python id": 1006, "finished": true, "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878640.135, "dur": 0.362,
    "args": {
      "Python parent id": 1004, "Python id": 1007, "finished": true, "Ev Idx": 3907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878640.391, "dur": 0.044,
    "args": {
      "Python parent id": 1007, "Python id": 1008, "finished": true, "Ev Idx": 3908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878640.851, "dur": 73.235,
    "args": {
      "Python parent id": 1004, "Python id": 1009, "finished": true, "Ev Idx": 3909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878641.500, "dur": 71.424,
    "args": {
      "Python parent id": 1009, "Python id": 1010, "finished": true, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878648.124, "dur": 16.951,
    "args": {
      "Python parent id": 1010, "Python id": 1011, "finished": true, "Ev Idx": 3911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878648.256, "dur": 0.112,
    "args": {
      "Python parent id": 1011, "Python id": 1012, "finished": true, "Ev Idx": 3912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878648.636, "dur": 16.359,
    "args": {
      "Python parent id": 1011, "Python id": 1013, "finished": true, "Ev Idx": 3913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878648.737, "dur": 7.524,
    "args": {
      "Python parent id": 1013, "Python id": 1014, "finished": true, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878672.556, "dur": 7.162,
    "args": {
      "Python parent id": 1010, "Python id": 1015, "finished": true, "Ev Idx": 3915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878688.540, "dur": 8.587,
    "args": {
      "Python parent id": 1010, "Python id": 1016, "finished": true, "Ev Idx": 3916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878716.441, "dur": 0.103,
    "args": {
      "Python parent id": 12, "Python id": 1017, "finished": true, "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878716.889, "dur": 0.157,
    "args": {
      "Python parent id": 12, "Python id": 1018, "finished": true, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878717.300, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 1019, "finished": true, "Ev Idx": 3919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878717.537, "dur": 9.193,
    "args": {
      "Python parent id": 12, "Python id": 1020, "finished": true, "Ev Idx": 3920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878727.741, "dur": 83.968,
    "args": {
      "Python parent id": 12, "Python id": 1021, "finished": true, "Ev Idx": 3921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878728.064, "dur": 0.432,
    "args": {
      "Python parent id": 1021, "Python id": 1022, "finished": true, "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878728.350, "dur": 0.080,
    "args": {
      "Python parent id": 1022, "Python id": 1023, "finished": true, "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878728.716, "dur": 0.362,
    "args": {
      "Python parent id": 1021, "Python id": 1024, "finished": true, "Ev Idx": 3924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878729.360, "dur": 0.241,
    "args": {
      "Python parent id": 1021, "Python id": 1025, "finished": true, "Ev Idx": 3925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878729.846, "dur": 1.280,
    "args": {
      "Python parent id": 1021, "Python id": 1026, "finished": true, "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878731.732, "dur": 79.877,
    "args": {
      "Python parent id": 1021, "Python id": 1027, "finished": true, "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878732.467, "dur": 78.999,
    "args": {
      "Python parent id": 1027, "Python id": 1028, "finished": true, "Ev Idx": 3928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878732.704, "dur": 78.702,
    "args": {
      "Python parent id": 1028, "Python id": 1029, "finished": true, "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878736.870, "dur": 73.483,
    "args": {
      "Python parent id": 1029, "Python id": 1030, "finished": true, "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878737.122, "dur": 73.098,
    "args": {
      "Python parent id": 1030, "Python id": 1031, "finished": true, "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878737.640, "dur": 72.496,
    "args": {
      "Python parent id": 1031, "Python id": 1032, "finished": true, "Ev Idx": 3932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878738.150, "dur": 0.329,
    "args": {
      "Python parent id": 1032, "Python id": 1033, "finished": true, "Ev Idx": 3933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878738.378, "dur": 0.044,
    "args": {
      "Python parent id": 1033, "Python id": 1034, "finished": true, "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878738.746, "dur": 0.330,
    "args": {
      "Python parent id": 1032, "Python id": 1035, "finished": true, "Ev Idx": 3935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878738.956, "dur": 0.076,
    "args": {
      "Python parent id": 1035, "Python id": 1036, "finished": true, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878739.441, "dur": 70.614,
    "args": {
      "Python parent id": 1032, "Python id": 1037, "finished": true, "Ev Idx": 3937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878740.022, "dur": 68.862,
    "args": {
      "Python parent id": 1037, "Python id": 1038, "finished": true, "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878746.487, "dur": 16.531,
    "args": {
      "Python parent id": 1038, "Python id": 1039, "finished": true, "Ev Idx": 3939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878746.622, "dur": 0.091,
    "args": {
      "Python parent id": 1039, "Python id": 1040, "finished": true, "Ev Idx": 3940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878746.928, "dur": 16.002,
    "args": {
      "Python parent id": 1039, "Python id": 1041, "finished": true, "Ev Idx": 3941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878747.039, "dur": 7.743,
    "args": {
      "Python parent id": 1041, "Python id": 1042, "finished": true, "Ev Idx": 3942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878769.983, "dur": 7.450,
    "args": {
      "Python parent id": 1038, "Python id": 1043, "finished": true, "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878786.107, "dur": 7.990,
    "args": {
      "Python parent id": 1038, "Python id": 1044, "finished": true, "Ev Idx": 3944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878812.298, "dur": 0.104,
    "args": {
      "Python parent id": 12, "Python id": 1045, "finished": true, "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878812.720, "dur": 0.143,
    "args": {
      "Python parent id": 12, "Python id": 1046, "finished": true, "Ev Idx": 3946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878813.113, "dur": 0.041,
    "args": {
      "Python parent id": 12, "Python id": 1047, "finished": true, "Ev Idx": 3947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878813.359, "dur": 9.303,
    "args": {
      "Python parent id": 12, "Python id": 1048, "finished": true, "Ev Idx": 3948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878823.712, "dur": 83.795,
    "args": {
      "Python parent id": 12, "Python id": 1049, "finished": true, "Ev Idx": 3949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878824.118, "dur": 0.448,
    "args": {
      "Python parent id": 1049, "Python id": 1050, "finished": true, "Ev Idx": 3950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878824.402, "dur": 0.102,
    "args": {
      "Python parent id": 1050, "Python id": 1051, "finished": true, "Ev Idx": 3951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878824.855, "dur": 0.342,
    "args": {
      "Python parent id": 1049, "Python id": 1052, "finished": true, "Ev Idx": 3952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878825.418, "dur": 0.228,
    "args": {
      "Python parent id": 1049, "Python id": 1053, "finished": true, "Ev Idx": 3953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878825.920, "dur": 1.302,
    "args": {
      "Python parent id": 1049, "Python id": 1054, "finished": true, "Ev Idx": 3954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878827.771, "dur": 79.661,
    "args": {
      "Python parent id": 1049, "Python id": 1055, "finished": true, "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878828.603, "dur": 78.681,
    "args": {
      "Python parent id": 1055, "Python id": 1056, "finished": true, "Ev Idx": 3956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878828.793, "dur": 78.435,
    "args": {
      "Python parent id": 1056, "Python id": 1057, "finished": true, "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878832.941, "dur": 73.233,
    "args": {
      "Python parent id": 1057, "Python id": 1058, "finished": true, "Ev Idx": 3958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878833.194, "dur": 72.922,
    "args": {
      "Python parent id": 1058, "Python id": 1059, "finished": true, "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878833.692, "dur": 72.229,
    "args": {
      "Python parent id": 1059, "Python id": 1060, "finished": true, "Ev Idx": 3960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878834.189, "dur": 0.305,
    "args": {
      "Python parent id": 1060, "Python id": 1061, "finished": true, "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878834.398, "dur": 0.038,
    "args": {
      "Python parent id": 1061, "Python id": 1062, "finished": true, "Ev Idx": 3962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878834.780, "dur": 0.340,
    "args": {
      "Python parent id": 1060, "Python id": 1063, "finished": true, "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878835.018, "dur": 0.048,
    "args": {
      "Python parent id": 1063, "Python id": 1064, "finished": true, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878835.451, "dur": 70.349,
    "args": {
      "Python parent id": 1060, "Python id": 1065, "finished": true, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878836.018, "dur": 68.647,
    "args": {
      "Python parent id": 1065, "Python id": 1066, "finished": true, "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878842.497, "dur": 16.106,
    "args": {
      "Python parent id": 1066, "Python id": 1067, "finished": true, "Ev Idx": 3967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878842.632, "dur": 0.091,
    "args": {
      "Python parent id": 1067, "Python id": 1068, "finished": true, "Ev Idx": 3968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878842.962, "dur": 15.567,
    "args": {
      "Python parent id": 1067, "Python id": 1069, "finished": true, "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878843.082, "dur": 6.427,
    "args": {
      "Python parent id": 1069, "Python id": 1070, "finished": true, "Ev Idx": 3970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878866.307, "dur": 6.749,
    "args": {
      "Python parent id": 1066, "Python id": 1071, "finished": true, "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878881.588, "dur": 7.885,
    "args": {
      "Python parent id": 1066, "Python id": 1072, "finished": true, "Ev Idx": 3972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740878908.229, "dur": 0.106,
    "args": {
      "Python parent id": 12, "Python id": 1073, "finished": true, "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878908.671, "dur": 0.153,
    "args": {
      "Python parent id": 12, "Python id": 1074, "finished": true, "Ev Idx": 3974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740878909.060, "dur": 0.043,
    "args": {
      "Python parent id": 12, "Python id": 1075, "finished": true, "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740878909.313, "dur": 9.313,
    "args": {
      "Python parent id": 12, "Python id": 1076, "finished": true, "Ev Idx": 3976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740878919.630, "dur": 85.698,
    "args": {
      "Python parent id": 12, "Python id": 1077, "finished": true, "Ev Idx": 3977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878919.972, "dur": 0.448,
    "args": {
      "Python parent id": 1077, "Python id": 1078, "finished": true, "Ev Idx": 3978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878920.261, "dur": 0.083,
    "args": {
      "Python parent id": 1078, "Python id": 1079, "finished": true, "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878920.652, "dur": 0.343,
    "args": {
      "Python parent id": 1077, "Python id": 1080, "finished": true, "Ev Idx": 3980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740878921.264, "dur": 0.292,
    "args": {
      "Python parent id": 1077, "Python id": 1081, "finished": true, "Ev Idx": 3981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740878921.803, "dur": 1.274,
    "args": {
      "Python parent id": 1077, "Python id": 1082, "finished": true, "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740878923.596, "dur": 81.653,
    "args": {
      "Python parent id": 1077, "Python id": 1083, "finished": true, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740878924.386, "dur": 80.716,
    "args": {
      "Python parent id": 1083, "Python id": 1084, "finished": true, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878924.612, "dur": 80.409,
    "args": {
      "Python parent id": 1084, "Python id": 1085, "finished": true, "Ev Idx": 3985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740878929.003, "dur": 74.957,
    "args": {
      "Python parent id": 1085, "Python id": 1086, "finished": true, "Ev Idx": 3986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740878929.253, "dur": 74.617,
    "args": {
      "Python parent id": 1086, "Python id": 1087, "finished": true, "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740878929.755, "dur": 74.045,
    "args": {
      "Python parent id": 1087, "Python id": 1088, "finished": true, "Ev Idx": 3988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878930.268, "dur": 0.339,
    "args": {
      "Python parent id": 1088, "Python id": 1089, "finished": true, "Ev Idx": 3989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878930.506, "dur": 0.044,
    "args": {
      "Python parent id": 1089, "Python id": 1090, "finished": true, "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740878930.864, "dur": 0.353,
    "args": {
      "Python parent id": 1088, "Python id": 1091, "finished": true, "Ev Idx": 3991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740878931.116, "dur": 0.045,
    "args": {
      "Python parent id": 1091, "Python id": 1092, "finished": true, "Ev Idx": 3992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878931.534, "dur": 72.152,
    "args": {
      "Python parent id": 1088, "Python id": 1093, "finished": true, "Ev Idx": 3993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740878932.162, "dur": 70.423,
    "args": {
      "Python parent id": 1093, "Python id": 1094, "finished": true, "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740878938.656, "dur": 17.167,
    "args": {
      "Python parent id": 1094, "Python id": 1095, "finished": true, "Ev Idx": 3995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740878938.769, "dur": 0.088,
    "args": {
      "Python parent id": 1095, "Python id": 1096, "finished": true, "Ev Idx": 3996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740878939.074, "dur": 16.659,
    "args": {
      "Python parent id": 1095, "Python id": 1097, "finished": true, "Ev Idx": 3997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740878939.220, "dur": 7.876,
    "args": {
      "Python parent id": 1097, "Python id": 1098, "finished": true, "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878964.212, "dur": 6.672,
    "args": {
      "Python parent id": 1094, "Python id": 1099, "finished": true, "Ev Idx": 3999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740878978.905, "dur": 8.136,
    "args": {
      "Python parent id": 1094, "Python id": 1100, "finished": true, "Ev Idx": 4000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879006.076, "dur": 0.104,
    "args": {
      "Python parent id": 12, "Python id": 1101, "finished": true, "Ev Idx": 4001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879006.501, "dur": 0.149,
    "args": {
      "Python parent id": 12, "Python id": 1102, "finished": true, "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879006.893, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 1103, "finished": true, "Ev Idx": 4003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879007.134, "dur": 9.499,
    "args": {
      "Python parent id": 12, "Python id": 1104, "finished": true, "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879017.620, "dur": 86.405,
    "args": {
      "Python parent id": 12, "Python id": 1105, "finished": true, "Ev Idx": 4005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879017.874, "dur": 0.500,
    "args": {
      "Python parent id": 1105, "Python id": 1106, "finished": true, "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879018.212, "dur": 0.091,
    "args": {
      "Python parent id": 1106, "Python id": 1107, "finished": true, "Ev Idx": 4007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879018.615, "dur": 0.295,
    "args": {
      "Python parent id": 1105, "Python id": 1108, "finished": true, "Ev Idx": 4008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879019.182, "dur": 0.219,
    "args": {
      "Python parent id": 1105, "Python id": 1109, "finished": true, "Ev Idx": 4009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879019.660, "dur": 1.149,
    "args": {
      "Python parent id": 1105, "Python id": 1110, "finished": true, "Ev Idx": 4010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879021.448, "dur": 82.507,
    "args": {
      "Python parent id": 1105, "Python id": 1111, "finished": true, "Ev Idx": 4011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879022.297, "dur": 81.514,
    "args": {
      "Python parent id": 1111, "Python id": 1112, "finished": true, "Ev Idx": 4012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879022.523, "dur": 81.230,
    "args": {
      "Python parent id": 1112, "Python id": 1113, "finished": true, "Ev Idx": 4013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879026.985, "dur": 75.685,
    "args": {
      "Python parent id": 1113, "Python id": 1114, "finished": true, "Ev Idx": 4014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879027.269, "dur": 75.300,
    "args": {
      "Python parent id": 1114, "Python id": 1115, "finished": true, "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879027.830, "dur": 74.619,
    "args": {
      "Python parent id": 1115, "Python id": 1116, "finished": true, "Ev Idx": 4016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879028.266, "dur": 0.319,
    "args": {
      "Python parent id": 1116, "Python id": 1117, "finished": true, "Ev Idx": 4017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879028.481, "dur": 0.043,
    "args": {
      "Python parent id": 1117, "Python id": 1118, "finished": true, "Ev Idx": 4018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879028.807, "dur": 0.321,
    "args": {
      "Python parent id": 1116, "Python id": 1119, "finished": true, "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879029.008, "dur": 0.075,
    "args": {
      "Python parent id": 1119, "Python id": 1120, "finished": true, "Ev Idx": 4020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879029.441, "dur": 72.930,
    "args": {
      "Python parent id": 1116, "Python id": 1121, "finished": true, "Ev Idx": 4021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879030.084, "dur": 71.132,
    "args": {
      "Python parent id": 1121, "Python id": 1122, "finished": true, "Ev Idx": 4022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879036.634, "dur": 16.550,
    "args": {
      "Python parent id": 1122, "Python id": 1123, "finished": true, "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879036.765, "dur": 0.088,
    "args": {
      "Python parent id": 1123, "Python id": 1124, "finished": true, "Ev Idx": 4024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879037.124, "dur": 15.962,
    "args": {
      "Python parent id": 1123, "Python id": 1125, "finished": true, "Ev Idx": 4025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879037.236, "dur": 8.020,
    "args": {
      "Python parent id": 1125, "Python id": 1126, "finished": true, "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879060.732, "dur": 7.293,
    "args": {
      "Python parent id": 1122, "Python id": 1127, "finished": true, "Ev Idx": 4027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879077.104, "dur": 8.986,
    "args": {
      "Python parent id": 1122, "Python id": 1128, "finished": true, "Ev Idx": 4028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879104.732, "dur": 0.112,
    "args": {
      "Python parent id": 12, "Python id": 1129, "finished": true, "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879105.163, "dur": 0.192,
    "args": {
      "Python parent id": 12, "Python id": 1130, "finished": true, "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879105.626, "dur": 0.055,
    "args": {
      "Python parent id": 12, "Python id": 1131, "finished": true, "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879105.885, "dur": 9.812,
    "args": {
      "Python parent id": 12, "Python id": 1132, "finished": true, "Ev Idx": 4032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879116.698, "dur": 87.238,
    "args": {
      "Python parent id": 12, "Python id": 1133, "finished": true, "Ev Idx": 4033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879117.065, "dur": 0.460,
    "args": {
      "Python parent id": 1133, "Python id": 1134, "finished": true, "Ev Idx": 4034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879117.367, "dur": 0.095,
    "args": {
      "Python parent id": 1134, "Python id": 1135, "finished": true, "Ev Idx": 4035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879117.762, "dur": 0.307,
    "args": {
      "Python parent id": 1133, "Python id": 1136, "finished": true, "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879118.319, "dur": 0.231,
    "args": {
      "Python parent id": 1133, "Python id": 1137, "finished": true, "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879118.806, "dur": 1.197,
    "args": {
      "Python parent id": 1133, "Python id": 1138, "finished": true, "Ev Idx": 4038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879120.688, "dur": 83.170,
    "args": {
      "Python parent id": 1133, "Python id": 1139, "finished": true, "Ev Idx": 4039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879121.405, "dur": 82.280,
    "args": {
      "Python parent id": 1139, "Python id": 1140, "finished": true, "Ev Idx": 4040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879121.592, "dur": 82.033,
    "args": {
      "Python parent id": 1140, "Python id": 1141, "finished": true, "Ev Idx": 4041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879127.978, "dur": 74.546,
    "args": {
      "Python parent id": 1141, "Python id": 1142, "finished": true, "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879128.270, "dur": 74.178,
    "args": {
      "Python parent id": 1142, "Python id": 1143, "finished": true, "Ev Idx": 4043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879128.822, "dur": 73.550,
    "args": {
      "Python parent id": 1143, "Python id": 1144, "finished": true, "Ev Idx": 4044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879129.315, "dur": 0.334,
    "args": {
      "Python parent id": 1144, "Python id": 1145, "finished": true, "Ev Idx": 4045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879129.541, "dur": 0.048,
    "args": {
      "Python parent id": 1145, "Python id": 1146, "finished": true, "Ev Idx": 4046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879129.890, "dur": 0.377,
    "args": {
      "Python parent id": 1144, "Python id": 1147, "finished": true, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879130.135, "dur": 0.088,
    "args": {
      "Python parent id": 1147, "Python id": 1148, "finished": true, "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879130.589, "dur": 71.681,
    "args": {
      "Python parent id": 1144, "Python id": 1149, "finished": true, "Ev Idx": 4049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879131.161, "dur": 69.927,
    "args": {
      "Python parent id": 1149, "Python id": 1150, "finished": true, "Ev Idx": 4050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879137.630, "dur": 17.208,
    "args": {
      "Python parent id": 1150, "Python id": 1151, "finished": true, "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879137.736, "dur": 0.090,
    "args": {
      "Python parent id": 1151, "Python id": 1152, "finished": true, "Ev Idx": 4052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879138.039, "dur": 16.725,
    "args": {
      "Python parent id": 1151, "Python id": 1153, "finished": true, "Ev Idx": 4053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879138.146, "dur": 7.030,
    "args": {
      "Python parent id": 1153, "Python id": 1154, "finished": true, "Ev Idx": 4054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879162.258, "dur": 6.916,
    "args": {
      "Python parent id": 1150, "Python id": 1155, "finished": true, "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879177.721, "dur": 8.335,
    "args": {
      "Python parent id": 1150, "Python id": 1156, "finished": true, "Ev Idx": 4056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879204.632, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 1157, "finished": true, "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879205.106, "dur": 0.194,
    "args": {
      "Python parent id": 12, "Python id": 1158, "finished": true, "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879205.530, "dur": 0.049,
    "args": {
      "Python parent id": 12, "Python id": 1159, "finished": true, "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879205.751, "dur": 9.365,
    "args": {
      "Python parent id": 12, "Python id": 1160, "finished": true, "Ev Idx": 4060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879216.104, "dur": 103.710,
    "args": {
      "Python parent id": 12, "Python id": 1161, "finished": true, "Ev Idx": 4061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879216.409, "dur": 0.511,
    "args": {
      "Python parent id": 1161, "Python id": 1162, "finished": true, "Ev Idx": 4062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879216.749, "dur": 0.124,
    "args": {
      "Python parent id": 1162, "Python id": 1163, "finished": true, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879217.203, "dur": 0.301,
    "args": {
      "Python parent id": 1161, "Python id": 1164, "finished": true, "Ev Idx": 4064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879217.756, "dur": 0.252,
    "args": {
      "Python parent id": 1161, "Python id": 1165, "finished": true, "Ev Idx": 4065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879218.223, "dur": 1.182,
    "args": {
      "Python parent id": 1161, "Python id": 1166, "finished": true, "Ev Idx": 4066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879220.027, "dur": 99.714,
    "args": {
      "Python parent id": 1161, "Python id": 1167, "finished": true, "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879220.673, "dur": 98.920,
    "args": {
      "Python parent id": 1167, "Python id": 1168, "finished": true, "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879220.894, "dur": 98.621,
    "args": {
      "Python parent id": 1168, "Python id": 1169, "finished": true, "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879225.000, "dur": 93.402,
    "args": {
      "Python parent id": 1169, "Python id": 1170, "finished": true, "Ev Idx": 4070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879225.275, "dur": 93.043,
    "args": {
      "Python parent id": 1170, "Python id": 1171, "finished": true, "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879225.754, "dur": 92.440,
    "args": {
      "Python parent id": 1171, "Python id": 1172, "finished": true, "Ev Idx": 4072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879226.187, "dur": 0.388,
    "args": {
      "Python parent id": 1172, "Python id": 1173, "finished": true, "Ev Idx": 4073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879226.468, "dur": 0.047,
    "args": {
      "Python parent id": 1173, "Python id": 1174, "finished": true, "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879226.835, "dur": 0.333,
    "args": {
      "Python parent id": 1172, "Python id": 1175, "finished": true, "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879227.070, "dur": 0.042,
    "args": {
      "Python parent id": 1175, "Python id": 1176, "finished": true, "Ev Idx": 4076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879227.493, "dur": 90.588,
    "args": {
      "Python parent id": 1172, "Python id": 1177, "finished": true, "Ev Idx": 4077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879228.073, "dur": 88.809,
    "args": {
      "Python parent id": 1177, "Python id": 1178, "finished": true, "Ev Idx": 4078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879234.545, "dur": 16.482,
    "args": {
      "Python parent id": 1178, "Python id": 1179, "finished": true, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879234.645, "dur": 0.090,
    "args": {
      "Python parent id": 1179, "Python id": 1180, "finished": true, "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879234.952, "dur": 15.993,
    "args": {
      "Python parent id": 1179, "Python id": 1181, "finished": true, "Ev Idx": 4081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879235.063, "dur": 6.808,
    "args": {
      "Python parent id": 1181, "Python id": 1182, "finished": true, "Ev Idx": 4082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879258.887, "dur": 19.558,
    "args": {
      "Python parent id": 1178, "Python id": 1183, "finished": true, "Ev Idx": 4083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879286.872, "dur": 9.132,
    "args": {
      "Python parent id": 1178, "Python id": 1184, "finished": true, "Ev Idx": 4084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879320.444, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 1185, "finished": true, "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879320.884, "dur": 0.118,
    "args": {
      "Python parent id": 12, "Python id": 1186, "finished": true, "Ev Idx": 4086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879321.248, "dur": 0.050,
    "args": {
      "Python parent id": 12, "Python id": 1187, "finished": true, "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879321.473, "dur": 9.454,
    "args": {
      "Python parent id": 12, "Python id": 1188, "finished": true, "Ev Idx": 4088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879332.016, "dur": 90.222,
    "args": {
      "Python parent id": 12, "Python id": 1189, "finished": true, "Ev Idx": 4089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879332.334, "dur": 0.472,
    "args": {
      "Python parent id": 1189, "Python id": 1190, "finished": true, "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879332.656, "dur": 0.086,
    "args": {
      "Python parent id": 1190, "Python id": 1191, "finished": true, "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879333.056, "dur": 0.323,
    "args": {
      "Python parent id": 1189, "Python id": 1192, "finished": true, "Ev Idx": 4092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879333.582, "dur": 0.260,
    "args": {
      "Python parent id": 1189, "Python id": 1193, "finished": true, "Ev Idx": 4093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879334.089, "dur": 1.111,
    "args": {
      "Python parent id": 1189, "Python id": 1194, "finished": true, "Ev Idx": 4094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879335.760, "dur": 86.402,
    "args": {
      "Python parent id": 1189, "Python id": 1195, "finished": true, "Ev Idx": 4095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879336.532, "dur": 85.439,
    "args": {
      "Python parent id": 1195, "Python id": 1196, "finished": true, "Ev Idx": 4096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879336.772, "dur": 85.146,
    "args": {
      "Python parent id": 1196, "Python id": 1197, "finished": true, "Ev Idx": 4097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879341.069, "dur": 79.717,
    "args": {
      "Python parent id": 1197, "Python id": 1198, "finished": true, "Ev Idx": 4098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879341.304, "dur": 79.396,
    "args": {
      "Python parent id": 1198, "Python id": 1199, "finished": true, "Ev Idx": 4099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879341.797, "dur": 78.836,
    "args": {
      "Python parent id": 1199, "Python id": 1200, "finished": true, "Ev Idx": 4100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879342.265, "dur": 0.383,
    "args": {
      "Python parent id": 1200, "Python id": 1201, "finished": true, "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879342.545, "dur": 0.047,
    "args": {
      "Python parent id": 1201, "Python id": 1202, "finished": true, "Ev Idx": 4102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879342.922, "dur": 0.335,
    "args": {
      "Python parent id": 1200, "Python id": 1203, "finished": true, "Ev Idx": 4103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879343.159, "dur": 0.044,
    "args": {
      "Python parent id": 1203, "Python id": 1204, "finished": true, "Ev Idx": 4104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879343.550, "dur": 76.970,
    "args": {
      "Python parent id": 1200, "Python id": 1205, "finished": true, "Ev Idx": 4105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879344.138, "dur": 75.181,
    "args": {
      "Python parent id": 1205, "Python id": 1206, "finished": true, "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879350.867, "dur": 18.278,
    "args": {
      "Python parent id": 1206, "Python id": 1207, "finished": true, "Ev Idx": 4107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879350.991, "dur": 0.087,
    "args": {
      "Python parent id": 1207, "Python id": 1208, "finished": true, "Ev Idx": 4108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879351.313, "dur": 17.753,
    "args": {
      "Python parent id": 1207, "Python id": 1209, "finished": true, "Ev Idx": 4109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879351.422, "dur": 8.283,
    "args": {
      "Python parent id": 1209, "Python id": 1210, "finished": true, "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879377.242, "dur": 6.864,
    "args": {
      "Python parent id": 1206, "Python id": 1211, "finished": true, "Ev Idx": 4111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879393.847, "dur": 8.419,
    "args": {
      "Python parent id": 1206, "Python id": 1212, "finished": true, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879422.903, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 1213, "finished": true, "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879423.290, "dur": 0.163,
    "args": {
      "Python parent id": 12, "Python id": 1214, "finished": true, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879423.699, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 1215, "finished": true, "Ev Idx": 4115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879423.907, "dur": 10.633,
    "args": {
      "Python parent id": 12, "Python id": 1216, "finished": true, "Ev Idx": 4116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879435.520, "dur": 84.935,
    "args": {
      "Python parent id": 12, "Python id": 1217, "finished": true, "Ev Idx": 4117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879435.806, "dur": 0.501,
    "args": {
      "Python parent id": 1217, "Python id": 1218, "finished": true, "Ev Idx": 4118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879436.153, "dur": 0.085,
    "args": {
      "Python parent id": 1218, "Python id": 1219, "finished": true, "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879436.566, "dur": 0.349,
    "args": {
      "Python parent id": 1217, "Python id": 1220, "finished": true, "Ev Idx": 4120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879437.125, "dur": 0.238,
    "args": {
      "Python parent id": 1217, "Python id": 1221, "finished": true, "Ev Idx": 4121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879437.593, "dur": 1.296,
    "args": {
      "Python parent id": 1217, "Python id": 1222, "finished": true, "Ev Idx": 4122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879439.482, "dur": 80.894,
    "args": {
      "Python parent id": 1217, "Python id": 1223, "finished": true, "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879440.260, "dur": 79.955,
    "args": {
      "Python parent id": 1223, "Python id": 1224, "finished": true, "Ev Idx": 4124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879440.486, "dur": 79.671,
    "args": {
      "Python parent id": 1224, "Python id": 1225, "finished": true, "Ev Idx": 4125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879444.618, "dur": 74.456,
    "args": {
      "Python parent id": 1225, "Python id": 1226, "finished": true, "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879444.900, "dur": 74.050,
    "args": {
      "Python parent id": 1226, "Python id": 1227, "finished": true, "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879445.361, "dur": 73.513,
    "args": {
      "Python parent id": 1227, "Python id": 1228, "finished": true, "Ev Idx": 4128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879445.841, "dur": 0.326,
    "args": {
      "Python parent id": 1228, "Python id": 1229, "finished": true, "Ev Idx": 4129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879446.058, "dur": 0.039,
    "args": {
      "Python parent id": 1229, "Python id": 1230, "finished": true, "Ev Idx": 4130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879446.392, "dur": 0.314,
    "args": {
      "Python parent id": 1228, "Python id": 1231, "finished": true, "Ev Idx": 4131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879446.610, "dur": 0.041,
    "args": {
      "Python parent id": 1231, "Python id": 1232, "finished": true, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879447.023, "dur": 71.776,
    "args": {
      "Python parent id": 1228, "Python id": 1233, "finished": true, "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879447.504, "dur": 70.075,
    "args": {
      "Python parent id": 1233, "Python id": 1234, "finished": true, "Ev Idx": 4134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879453.923, "dur": 17.116,
    "args": {
      "Python parent id": 1234, "Python id": 1235, "finished": true, "Ev Idx": 4135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879454.019, "dur": 0.085,
    "args": {
      "Python parent id": 1235, "Python id": 1236, "finished": true, "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879454.331, "dur": 16.617,
    "args": {
      "Python parent id": 1235, "Python id": 1237, "finished": true, "Ev Idx": 4137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879454.477, "dur": 6.604,
    "args": {
      "Python parent id": 1237, "Python id": 1238, "finished": true, "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879479.043, "dur": 6.978,
    "args": {
      "Python parent id": 1234, "Python id": 1239, "finished": true, "Ev Idx": 4139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879494.797, "dur": 8.209,
    "args": {
      "Python parent id": 1234, "Python id": 1240, "finished": true, "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879521.102, "dur": 0.105,
    "args": {
      "Python parent id": 12, "Python id": 1241, "finished": true, "Ev Idx": 4141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879521.524, "dur": 0.160,
    "args": {
      "Python parent id": 12, "Python id": 1242, "finished": true, "Ev Idx": 4142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879521.908, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 1243, "finished": true, "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879522.108, "dur": 8.638,
    "args": {
      "Python parent id": 12, "Python id": 1244, "finished": true, "Ev Idx": 4144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879531.702, "dur": 85.732,
    "args": {
      "Python parent id": 12, "Python id": 1245, "finished": true, "Ev Idx": 4145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879531.999, "dur": 0.527,
    "args": {
      "Python parent id": 1245, "Python id": 1246, "finished": true, "Ev Idx": 4146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879532.381, "dur": 0.080,
    "args": {
      "Python parent id": 1246, "Python id": 1247, "finished": true, "Ev Idx": 4147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879532.811, "dur": 0.841,
    "args": {
      "Python parent id": 1245, "Python id": 1248, "finished": true, "Ev Idx": 4148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879533.893, "dur": 0.232,
    "args": {
      "Python parent id": 1245, "Python id": 1249, "finished": true, "Ev Idx": 4149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879534.349, "dur": 1.236,
    "args": {
      "Python parent id": 1245, "Python id": 1250, "finished": true, "Ev Idx": 4150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879536.165, "dur": 81.198,
    "args": {
      "Python parent id": 1245, "Python id": 1251, "finished": true, "Ev Idx": 4151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879536.848, "dur": 80.366,
    "args": {
      "Python parent id": 1251, "Python id": 1252, "finished": true, "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879537.047, "dur": 80.105,
    "args": {
      "Python parent id": 1252, "Python id": 1253, "finished": true, "Ev Idx": 4153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879541.101, "dur": 74.875,
    "args": {
      "Python parent id": 1253, "Python id": 1254, "finished": true, "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879541.400, "dur": 74.494,
    "args": {
      "Python parent id": 1254, "Python id": 1255, "finished": true, "Ev Idx": 4155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879541.908, "dur": 73.916,
    "args": {
      "Python parent id": 1255, "Python id": 1256, "finished": true, "Ev Idx": 4156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879542.440, "dur": 0.343,
    "args": {
      "Python parent id": 1256, "Python id": 1257, "finished": true, "Ev Idx": 4157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879542.685, "dur": 0.042,
    "args": {
      "Python parent id": 1257, "Python id": 1258, "finished": true, "Ev Idx": 4158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879543.073, "dur": 0.330,
    "args": {
      "Python parent id": 1256, "Python id": 1259, "finished": true, "Ev Idx": 4159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879543.315, "dur": 0.038,
    "args": {
      "Python parent id": 1259, "Python id": 1260, "finished": true, "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879543.677, "dur": 72.017,
    "args": {
      "Python parent id": 1256, "Python id": 1261, "finished": true, "Ev Idx": 4161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879544.235, "dur": 70.255,
    "args": {
      "Python parent id": 1261, "Python id": 1262, "finished": true, "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879550.792, "dur": 17.280,
    "args": {
      "Python parent id": 1262, "Python id": 1263, "finished": true, "Ev Idx": 4163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879550.891, "dur": 0.089,
    "args": {
      "Python parent id": 1263, "Python id": 1264, "finished": true, "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879551.243, "dur": 16.755,
    "args": {
      "Python parent id": 1263, "Python id": 1265, "finished": true, "Ev Idx": 4165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879551.351, "dur": 7.621,
    "args": {
      "Python parent id": 1265, "Python id": 1266, "finished": true, "Ev Idx": 4166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879575.420, "dur": 7.273,
    "args": {
      "Python parent id": 1262, "Python id": 1267, "finished": true, "Ev Idx": 4167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879591.132, "dur": 7.607,
    "args": {
      "Python parent id": 1262, "Python id": 1268, "finished": true, "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879618.136, "dur": 0.110,
    "args": {
      "Python parent id": 12, "Python id": 1269, "finished": true, "Ev Idx": 4169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879618.581, "dur": 0.155,
    "args": {
      "Python parent id": 12, "Python id": 1270, "finished": true, "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879618.952, "dur": 0.041,
    "args": {
      "Python parent id": 12, "Python id": 1271, "finished": true, "Ev Idx": 4171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879619.177, "dur": 9.234,
    "args": {
      "Python parent id": 12, "Python id": 1272, "finished": true, "Ev Idx": 4172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879629.360, "dur": 85.079,
    "args": {
      "Python parent id": 12, "Python id": 1273, "finished": true, "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879629.677, "dur": 0.507,
    "args": {
      "Python parent id": 1273, "Python id": 1274, "finished": true, "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879630.041, "dur": 0.080,
    "args": {
      "Python parent id": 1274, "Python id": 1275, "finished": true, "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879630.464, "dur": 0.315,
    "args": {
      "Python parent id": 1273, "Python id": 1276, "finished": true, "Ev Idx": 4176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879631.031, "dur": 0.257,
    "args": {
      "Python parent id": 1273, "Python id": 1277, "finished": true, "Ev Idx": 4177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879631.496, "dur": 1.264,
    "args": {
      "Python parent id": 1273, "Python id": 1278, "finished": true, "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879633.344, "dur": 81.001,
    "args": {
      "Python parent id": 1273, "Python id": 1279, "finished": true, "Ev Idx": 4179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879633.994, "dur": 80.141,
    "args": {
      "Python parent id": 1279, "Python id": 1280, "finished": true, "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879634.170, "dur": 79.909,
    "args": {
      "Python parent id": 1280, "Python id": 1281, "finished": true, "Ev Idx": 4181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879638.273, "dur": 74.759,
    "args": {
      "Python parent id": 1281, "Python id": 1282, "finished": true, "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879638.518, "dur": 74.376,
    "args": {
      "Python parent id": 1282, "Python id": 1283, "finished": true, "Ev Idx": 4183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879638.961, "dur": 73.855,
    "args": {
      "Python parent id": 1283, "Python id": 1284, "finished": true, "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879639.362, "dur": 0.383,
    "args": {
      "Python parent id": 1284, "Python id": 1285, "finished": true, "Ev Idx": 4185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879639.629, "dur": 0.049,
    "args": {
      "Python parent id": 1285, "Python id": 1286, "finished": true, "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879640.024, "dur": 0.357,
    "args": {
      "Python parent id": 1284, "Python id": 1287, "finished": true, "Ev Idx": 4187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879640.264, "dur": 0.074,
    "args": {
      "Python parent id": 1287, "Python id": 1288, "finished": true, "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879640.741, "dur": 71.997,
    "args": {
      "Python parent id": 1284, "Python id": 1289, "finished": true, "Ev Idx": 4189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879641.296, "dur": 70.285,
    "args": {
      "Python parent id": 1289, "Python id": 1290, "finished": true, "Ev Idx": 4190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879647.776, "dur": 16.820,
    "args": {
      "Python parent id": 1290, "Python id": 1291, "finished": true, "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879647.869, "dur": 0.092,
    "args": {
      "Python parent id": 1291, "Python id": 1292, "finished": true, "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879648.197, "dur": 16.321,
    "args": {
      "Python parent id": 1291, "Python id": 1293, "finished": true, "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879648.332, "dur": 7.796,
    "args": {
      "Python parent id": 1293, "Python id": 1294, "finished": true, "Ev Idx": 4194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879671.942, "dur": 6.498,
    "args": {
      "Python parent id": 1290, "Python id": 1295, "finished": true, "Ev Idx": 4195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879687.613, "dur": 8.805,
    "args": {
      "Python parent id": 1290, "Python id": 1296, "finished": true, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879715.077, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 1297, "finished": true, "Ev Idx": 4197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879715.538, "dur": 0.145,
    "args": {
      "Python parent id": 12, "Python id": 1298, "finished": true, "Ev Idx": 4198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879715.928, "dur": 0.045,
    "args": {
      "Python parent id": 12, "Python id": 1299, "finished": true, "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879716.133, "dur": 9.416,
    "args": {
      "Python parent id": 12, "Python id": 1300, "finished": true, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879726.480, "dur": 85.399,
    "args": {
      "Python parent id": 12, "Python id": 1301, "finished": true, "Ev Idx": 4201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879726.794, "dur": 0.507,
    "args": {
      "Python parent id": 1301, "Python id": 1302, "finished": true, "Ev Idx": 4202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879727.145, "dur": 0.089,
    "args": {
      "Python parent id": 1302, "Python id": 1303, "finished": true, "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879727.544, "dur": 0.311,
    "args": {
      "Python parent id": 1301, "Python id": 1304, "finished": true, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879728.105, "dur": 0.260,
    "args": {
      "Python parent id": 1301, "Python id": 1305, "finished": true, "Ev Idx": 4205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879728.607, "dur": 1.274,
    "args": {
      "Python parent id": 1301, "Python id": 1306, "finished": true, "Ev Idx": 4206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879730.389, "dur": 81.419,
    "args": {
      "Python parent id": 1301, "Python id": 1307, "finished": true, "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879731.030, "dur": 80.604,
    "args": {
      "Python parent id": 1307, "Python id": 1308, "finished": true, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879731.201, "dur": 80.350,
    "args": {
      "Python parent id": 1308, "Python id": 1309, "finished": true, "Ev Idx": 4209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879735.180, "dur": 75.384,
    "args": {
      "Python parent id": 1309, "Python id": 1310, "finished": true, "Ev Idx": 4210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879735.480, "dur": 74.986,
    "args": {
      "Python parent id": 1310, "Python id": 1311, "finished": true, "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879735.934, "dur": 74.454,
    "args": {
      "Python parent id": 1311, "Python id": 1312, "finished": true, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879736.407, "dur": 0.362,
    "args": {
      "Python parent id": 1312, "Python id": 1313, "finished": true, "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879736.667, "dur": 0.047,
    "args": {
      "Python parent id": 1313, "Python id": 1314, "finished": true, "Ev Idx": 4214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879737.017, "dur": 0.361,
    "args": {
      "Python parent id": 1312, "Python id": 1315, "finished": true, "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879737.257, "dur": 0.076,
    "args": {
      "Python parent id": 1315, "Python id": 1316, "finished": true, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879737.687, "dur": 72.615,
    "args": {
      "Python parent id": 1312, "Python id": 1317, "finished": true, "Ev Idx": 4217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879738.218, "dur": 70.922,
    "args": {
      "Python parent id": 1317, "Python id": 1318, "finished": true, "Ev Idx": 4218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879744.646, "dur": 17.746,
    "args": {
      "Python parent id": 1318, "Python id": 1319, "finished": true, "Ev Idx": 4219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879744.793, "dur": 0.114,
    "args": {
      "Python parent id": 1319, "Python id": 1320, "finished": true, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879745.118, "dur": 17.178,
    "args": {
      "Python parent id": 1319, "Python id": 1321, "finished": true, "Ev Idx": 4221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879745.231, "dur": 7.386,
    "args": {
      "Python parent id": 1321, "Python id": 1322, "finished": true, "Ev Idx": 4222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879769.410, "dur": 7.077,
    "args": {
      "Python parent id": 1318, "Python id": 1323, "finished": true, "Ev Idx": 4223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879785.344, "dur": 8.291,
    "args": {
      "Python parent id": 1318, "Python id": 1324, "finished": true, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879812.528, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 1325, "finished": true, "Ev Idx": 4225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879812.969, "dur": 0.166,
    "args": {
      "Python parent id": 12, "Python id": 1326, "finished": true, "Ev Idx": 4226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879813.377, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 1327, "finished": true, "Ev Idx": 4227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879813.594, "dur": 9.404,
    "args": {
      "Python parent id": 12, "Python id": 1328, "finished": true, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879823.954, "dur": 80.994,
    "args": {
      "Python parent id": 12, "Python id": 1329, "finished": true, "Ev Idx": 4229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879824.266, "dur": 0.451,
    "args": {
      "Python parent id": 1329, "Python id": 1330, "finished": true, "Ev Idx": 4230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879824.571, "dur": 0.084,
    "args": {
      "Python parent id": 1330, "Python id": 1331, "finished": true, "Ev Idx": 4231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879824.961, "dur": 0.294,
    "args": {
      "Python parent id": 1329, "Python id": 1332, "finished": true, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879825.477, "dur": 0.238,
    "args": {
      "Python parent id": 1329, "Python id": 1333, "finished": true, "Ev Idx": 4233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879825.960, "dur": 1.214,
    "args": {
      "Python parent id": 1329, "Python id": 1334, "finished": true, "Ev Idx": 4234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879827.692, "dur": 77.181,
    "args": {
      "Python parent id": 1329, "Python id": 1335, "finished": true, "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879828.320, "dur": 76.403,
    "args": {
      "Python parent id": 1335, "Python id": 1336, "finished": true, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879828.488, "dur": 76.176,
    "args": {
      "Python parent id": 1336, "Python id": 1337, "finished": true, "Ev Idx": 4237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879832.624, "dur": 71.062,
    "args": {
      "Python parent id": 1337, "Python id": 1338, "finished": true, "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879832.882, "dur": 70.647,
    "args": {
      "Python parent id": 1338, "Python id": 1339, "finished": true, "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879833.354, "dur": 70.106,
    "args": {
      "Python parent id": 1339, "Python id": 1340, "finished": true, "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879833.804, "dur": 0.301,
    "args": {
      "Python parent id": 1340, "Python id": 1341, "finished": true, "Ev Idx": 4241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879834.006, "dur": 0.037,
    "args": {
      "Python parent id": 1341, "Python id": 1342, "finished": true, "Ev Idx": 4242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879834.345, "dur": 0.375,
    "args": {
      "Python parent id": 1340, "Python id": 1343, "finished": true, "Ev Idx": 4243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879834.587, "dur": 0.088,
    "args": {
      "Python parent id": 1343, "Python id": 1344, "finished": true, "Ev Idx": 4244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879835.102, "dur": 68.243,
    "args": {
      "Python parent id": 1340, "Python id": 1345, "finished": true, "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879835.694, "dur": 66.514,
    "args": {
      "Python parent id": 1345, "Python id": 1346, "finished": true, "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879842.206, "dur": 15.009,
    "args": {
      "Python parent id": 1346, "Python id": 1347, "finished": true, "Ev Idx": 4247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879842.297, "dur": 0.093,
    "args": {
      "Python parent id": 1347, "Python id": 1348, "finished": true, "Ev Idx": 4248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879842.600, "dur": 14.540,
    "args": {
      "Python parent id": 1347, "Python id": 1349, "finished": true, "Ev Idx": 4249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879842.709, "dur": 6.679,
    "args": {
      "Python parent id": 1349, "Python id": 1350, "finished": true, "Ev Idx": 4250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879863.905, "dur": 7.445,
    "args": {
      "Python parent id": 1346, "Python id": 1351, "finished": true, "Ev Idx": 4251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879879.781, "dur": 7.820,
    "args": {
      "Python parent id": 1346, "Python id": 1352, "finished": true, "Ev Idx": 4252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740879905.606, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 1353, "finished": true, "Ev Idx": 4253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879906.036, "dur": 0.151,
    "args": {
      "Python parent id": 12, "Python id": 1354, "finished": true, "Ev Idx": 4254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740879906.432, "dur": 0.044,
    "args": {
      "Python parent id": 12, "Python id": 1355, "finished": true, "Ev Idx": 4255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740879906.642, "dur": 9.000,
    "args": {
      "Python parent id": 12, "Python id": 1356, "finished": true, "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740879916.577, "dur": 83.027,
    "args": {
      "Python parent id": 12, "Python id": 1357, "finished": true, "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879916.917, "dur": 0.441,
    "args": {
      "Python parent id": 1357, "Python id": 1358, "finished": true, "Ev Idx": 4258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879917.209, "dur": 0.085,
    "args": {
      "Python parent id": 1358, "Python id": 1359, "finished": true, "Ev Idx": 4259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879917.688, "dur": 0.282,
    "args": {
      "Python parent id": 1357, "Python id": 1360, "finished": true, "Ev Idx": 4260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740879918.162, "dur": 0.248,
    "args": {
      "Python parent id": 1357, "Python id": 1361, "finished": true, "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740879918.634, "dur": 1.151,
    "args": {
      "Python parent id": 1357, "Python id": 1362, "finished": true, "Ev Idx": 4262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740879920.264, "dur": 79.267,
    "args": {
      "Python parent id": 1357, "Python id": 1363, "finished": true, "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740879920.908, "dur": 78.471,
    "args": {
      "Python parent id": 1363, "Python id": 1364, "finished": true, "Ev Idx": 4264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879921.105, "dur": 78.217,
    "args": {
      "Python parent id": 1364, "Python id": 1365, "finished": true, "Ev Idx": 4265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740879925.245, "dur": 73.070,
    "args": {
      "Python parent id": 1365, "Python id": 1366, "finished": true, "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740879925.539, "dur": 72.651,
    "args": {
      "Python parent id": 1366, "Python id": 1367, "finished": true, "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740879926.023, "dur": 72.088,
    "args": {
      "Python parent id": 1367, "Python id": 1368, "finished": true, "Ev Idx": 4268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879926.472, "dur": 0.345,
    "args": {
      "Python parent id": 1368, "Python id": 1369, "finished": true, "Ev Idx": 4269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879926.711, "dur": 0.044,
    "args": {
      "Python parent id": 1369, "Python id": 1370, "finished": true, "Ev Idx": 4270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740879927.085, "dur": 0.305,
    "args": {
      "Python parent id": 1368, "Python id": 1371, "finished": true, "Ev Idx": 4271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740879927.294, "dur": 0.041,
    "args": {
      "Python parent id": 1371, "Python id": 1372, "finished": true, "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879927.681, "dur": 70.351,
    "args": {
      "Python parent id": 1368, "Python id": 1373, "finished": true, "Ev Idx": 4273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740879928.248, "dur": 68.650,
    "args": {
      "Python parent id": 1373, "Python id": 1374, "finished": true, "Ev Idx": 4274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740879934.696, "dur": 16.889,
    "args": {
      "Python parent id": 1374, "Python id": 1375, "finished": true, "Ev Idx": 4275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740879934.790, "dur": 0.089,
    "args": {
      "Python parent id": 1375, "Python id": 1376, "finished": true, "Ev Idx": 4276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740879935.156, "dur": 16.332,
    "args": {
      "Python parent id": 1375, "Python id": 1377, "finished": true, "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740879935.255, "dur": 6.986,
    "args": {
      "Python parent id": 1377, "Python id": 1378, "finished": true, "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879959.125, "dur": 6.736,
    "args": {
      "Python parent id": 1374, "Python id": 1379, "finished": true, "Ev Idx": 4279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740879974.506, "dur": 7.669,
    "args": {
      "Python parent id": 1374, "Python id": 1380, "finished": true, "Ev Idx": 4280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740880000.287, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 1381, "finished": true, "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880000.718, "dur": 0.127,
    "args": {
      "Python parent id": 12, "Python id": 1382, "finished": true, "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880001.110, "dur": 0.042,
    "args": {
      "Python parent id": 12, "Python id": 1383, "finished": true, "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740880001.325, "dur": 9.155,
    "args": {
      "Python parent id": 12, "Python id": 1384, "finished": true, "Ev Idx": 4284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740880011.497, "dur": 82.848,
    "args": {
      "Python parent id": 12, "Python id": 1385, "finished": true, "Ev Idx": 4285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880011.878, "dur": 0.467,
    "args": {
      "Python parent id": 1385, "Python id": 1386, "finished": true, "Ev Idx": 4286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880012.198, "dur": 0.084,
    "args": {
      "Python parent id": 1386, "Python id": 1387, "finished": true, "Ev Idx": 4287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740880012.602, "dur": 0.278,
    "args": {
      "Python parent id": 1385, "Python id": 1388, "finished": true, "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740880013.087, "dur": 0.247,
    "args": {
      "Python parent id": 1385, "Python id": 1389, "finished": true, "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740880013.554, "dur": 1.149,
    "args": {
      "Python parent id": 1385, "Python id": 1390, "finished": true, "Ev Idx": 4290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740880015.206, "dur": 79.067,
    "args": {
      "Python parent id": 1385, "Python id": 1391, "finished": true, "Ev Idx": 4291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740880015.784, "dur": 78.332,
    "args": {
      "Python parent id": 1391, "Python id": 1392, "finished": true, "Ev Idx": 4292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740880016.020, "dur": 78.043,
    "args": {
      "Python parent id": 1392, "Python id": 1393, "finished": true, "Ev Idx": 4293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740880020.070, "dur": 72.961,
    "args": {
      "Python parent id": 1393, "Python id": 1394, "finished": true, "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740880020.352, "dur": 72.563,
    "args": {
      "Python parent id": 1394, "Python id": 1395, "finished": true, "Ev Idx": 4295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740880020.798, "dur": 72.030,
    "args": {
      "Python parent id": 1395, "Python id": 1396, "finished": true, "Ev Idx": 4296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880021.251, "dur": 0.350,
    "args": {
      "Python parent id": 1396, "Python id": 1397, "finished": true, "Ev Idx": 4297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880021.498, "dur": 0.041,
    "args": {
      "Python parent id": 1397, "Python id": 1398, "finished": true, "Ev Idx": 4298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880021.857, "dur": 0.360,
    "args": {
      "Python parent id": 1396, "Python id": 1399, "finished": true, "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880022.126, "dur": 0.038,
    "args": {
      "Python parent id": 1399, "Python id": 1400, "finished": true, "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740880022.507, "dur": 70.243,
    "args": {
      "Python parent id": 1396, "Python id": 1401, "finished": true, "Ev Idx": 4301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740880023.057, "dur": 68.553,
    "args": {
      "Python parent id": 1401, "Python id": 1402, "finished": true, "Ev Idx": 4302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740880029.825, "dur": 16.507,
    "args": {
      "Python parent id": 1402, "Python id": 1403, "finished": true, "Ev Idx": 4303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740880029.925, "dur": 0.094,
    "args": {
      "Python parent id": 1403, "Python id": 1404, "finished": true, "Ev Idx": 4304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740880030.270, "dur": 15.946,
    "args": {
      "Python parent id": 1403, "Python id": 1405, "finished": true, "Ev Idx": 4305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740880030.418, "dur": 7.144,
    "args": {
      "Python parent id": 1405, "Python id": 1406, "finished": true, "Ev Idx": 4306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740880053.488, "dur": 6.625,
    "args": {
      "Python parent id": 1402, "Python id": 1407, "finished": true, "Ev Idx": 4307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740880068.392, "dur": 7.466,
    "args": {
      "Python parent id": 1402, "Python id": 1408, "finished": true, "Ev Idx": 4308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740880095.011, "dur": 0.113,
    "args": {
      "Python parent id": 12, "Python id": 1409, "finished": true, "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880095.471, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 1410, "finished": true, "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880095.806, "dur": 0.052,
    "args": {
      "Python parent id": 12, "Python id": 1411, "finished": true, "Ev Idx": 4311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740880096.054, "dur": 9.404,
    "args": {
      "Python parent id": 12, "Python id": 1412, "finished": true, "Ev Idx": 4312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740880106.397, "dur": 85.422,
    "args": {
      "Python parent id": 12, "Python id": 1413, "finished": true, "Ev Idx": 4313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880106.789, "dur": 0.518,
    "args": {
      "Python parent id": 1413, "Python id": 1414, "finished": true, "Ev Idx": 4314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880107.150, "dur": 0.093,
    "args": {
      "Python parent id": 1414, "Python id": 1415, "finished": true, "Ev Idx": 4315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740880107.507, "dur": 0.310,
    "args": {
      "Python parent id": 1413, "Python id": 1416, "finished": true, "Ev Idx": 4316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740880108.047, "dur": 0.234,
    "args": {
      "Python parent id": 1413, "Python id": 1417, "finished": true, "Ev Idx": 4317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740880108.494, "dur": 1.174,
    "args": {
      "Python parent id": 1413, "Python id": 1418, "finished": true, "Ev Idx": 4318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740880110.200, "dur": 81.537,
    "args": {
      "Python parent id": 1413, "Python id": 1419, "finished": true, "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740880111.005, "dur": 80.577,
    "args": {
      "Python parent id": 1419, "Python id": 1420, "finished": true, "Ev Idx": 4320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740880111.258, "dur": 80.249,
    "args": {
      "Python parent id": 1420, "Python id": 1421, "finished": true, "Ev Idx": 4321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740880115.556, "dur": 74.881,
    "args": {
      "Python parent id": 1421, "Python id": 1422, "finished": true, "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740880115.834, "dur": 74.441,
    "args": {
      "Python parent id": 1422, "Python id": 1423, "finished": true, "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740880116.307, "dur": 73.899,
    "args": {
      "Python parent id": 1423, "Python id": 1424, "finished": true, "Ev Idx": 4324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880116.746, "dur": 0.361,
    "args": {
      "Python parent id": 1424, "Python id": 1425, "finished": true, "Ev Idx": 4325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880117.004, "dur": 0.044,
    "args": {
      "Python parent id": 1425, "Python id": 1426, "finished": true, "Ev Idx": 4326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740880117.346, "dur": 0.290,
    "args": {
      "Python parent id": 1424, "Python id": 1427, "finished": true, "Ev Idx": 4327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740880117.542, "dur": 0.042,
    "args": {
      "Python parent id": 1427, "Python id": 1428, "finished": true, "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740880117.913, "dur": 72.189,
    "args": {
      "Python parent id": 1424, "Python id": 1429, "finished": true, "Ev Idx": 4329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740880118.444, "dur": 70.556,
    "args": {
      "Python parent id": 1429, "Python id": 1430, "finished": true, "Ev Idx": 4330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740880125.008, "dur": 16.199,
    "args": {
      "Python parent id": 1430, "Python id": 1431, "finished": true, "Ev Idx": 4331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740880125.158, "dur": 0.081,
    "args": {
      "Python parent id": 1431, "Python id": 1432, "finished": true, "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740880125.522, "dur": 15.593,
    "args": {
      "Python parent id": 1431, "Python id": 1433, "finished": true, "Ev Idx": 4333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740880125.624, "dur": 7.163,
    "args": {
      "Python parent id": 1433, "Python id": 1434, "finished": true, "Ev Idx": 4334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740880148.565, "dur": 7.407,
    "args": {
      "Python parent id": 1430, "Python id": 1435, "finished": true, "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740880164.876, "dur": 8.057,
    "args": {
      "Python parent id": 1430, "Python id": 1436, "finished": true, "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1030): synchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740880197.905, "dur": 20681.696,
    "args": {
      "Python parent id": 12, "Python id": 1437, "finished": true, "Ev Idx": 4337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740880201.288, "dur": 6.755,
    "args": {
      "Python parent id": 1437, "Python id": 1438, "finished": true, "Ev Idx": 4338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740880202.483, "dur": 5.285,
    "args": {
      "Python parent id": 1438, "Python id": 1439, "finished": true, "Ev Idx": 4339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740880207.128, "dur": 0.481,
    "args": {
      "Python parent id": 1439, "Python id": 1440, "finished": true, "Ev Idx": 4340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(490): __init__", "pid": 49625, "tid": 49625,
    "ts": 7253740880212.440, "dur": 65.840,
    "args": {
      "Python parent id": 1437, "Python id": 1441, "finished": true, "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740880215.161, "dur": 61.947,
    "args": {
      "Python parent id": 1441, "Python id": 1442, "finished": true, "Ev Idx": 4342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880215.608, "dur": 0.182,
    "args": {
      "Python parent id": 1442, "Python id": 1443, "finished": true, "Ev Idx": 4343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880216.693, "dur": 0.090,
    "args": {
      "Python parent id": 1442, "Python id": 1444, "finished": true, "Ev Idx": 4344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880217.897, "dur": 0.055,
    "args": {
      "Python parent id": 1442, "Python id": 1445, "finished": true, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740880221.443, "dur": 0.135,
    "args": {
      "Python parent id": 1442, "Python id": 1446, "finished": true, "Ev Idx": 4346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880222.394, "dur": 0.162,
    "args": {
      "Python parent id": 1442, "Python id": 1447, "finished": true, "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740880225.912, "dur": 51.089,
    "args": {
      "Python parent id": 1442, "Python id": 1448, "finished": true, "Ev Idx": 4348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880227.418, "dur": 0.048,
    "args": {
      "Python parent id": 1448, "Python id": 1449, "finished": true, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880227.923, "dur": 0.047,
    "args": {
      "Python parent id": 1448, "Python id": 1450, "finished": true, "Ev Idx": 4350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880228.534, "dur": 0.103,
    "args": {
      "Python parent id": 1448, "Python id": 1451, "finished": true, "Ev Idx": 4351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740880229.754, "dur": 0.047,
    "args": {
      "Python parent id": 1448, "Python id": 1452, "finished": true, "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740880231.708, "dur": 45.185,
    "args": {
      "Python parent id": 1448, "Python id": 1453, "finished": true, "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740880235.105, "dur": 41.521,
    "args": {
      "Python parent id": 1453, "Python id": 1454, "finished": true, "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740880238.121, "dur": 27.523,
    "args": {
      "Python parent id": 1454, "Python id": 1455, "finished": true, "Ev Idx": 4355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740880239.911, "dur": 25.584,
    "args": {
      "Python parent id": 1455, "Python id": 1456, "finished": true, "Ev Idx": 4356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740880243.746, "dur": 1.731,
    "args": {
      "Python parent id": 1456, "Python id": 1457, "finished": true, "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740880244.684, "dur": 0.701,
    "args": {
      "Python parent id": 1457, "Python id": 1458, "finished": true, "Ev Idx": 4358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740880246.700, "dur": 16.964,
    "args": {
      "Python parent id": 1456, "Python id": 1459, "finished": true, "Ev Idx": 4359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740880248.839, "dur": 14.518,
    "args": {
      "Python parent id": 1459, "Python id": 1460, "finished": true, "Ev Idx": 4360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740880250.368, "dur": 12.863,
    "args": {
      "Python parent id": 1460, "Python id": 1461, "finished": true, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740880251.650, "dur": 10.378,
    "args": {
      "Python parent id": 1461, "Python id": 1462, "finished": true, "Ev Idx": 4362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740880252.982, "dur": 5.847,
    "args": {
      "Python parent id": 1462, "Python id": 1463, "finished": true, "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740880253.316, "dur": 0.076,
    "args": {
      "Python parent id": 1463, "Python id": 1464, "finished": true, "Ev Idx": 4364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740880258.064, "dur": 0.625,
    "args": {
      "Python parent id": 1463, "Python id": 1465, "finished": true, "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740880264.865, "dur": 0.524,
    "args": {
      "Python parent id": 1456, "Python id": 1466, "finished": true, "Ev Idx": 4366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740880269.249, "dur": 0.345,
    "args": {
      "Python parent id": 1454, "Python id": 1467, "finished": true, "Ev Idx": 4367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740880271.218, "dur": 5.312,
    "args": {
      "Python parent id": 1454, "Python id": 1468, "finished": true, "Ev Idx": 4368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740880272.271, "dur": 4.103,
    "args": {
      "Python parent id": 1468, "Python id": 1469, "finished": true, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740880272.764, "dur": 1.062,
    "args": {
      "Python parent id": 1469, "Python id": 1470, "finished": true, "Ev Idx": 4370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740880273.228, "dur": 0.516,
    "args": {
      "Python parent id": 1470, "Python id": 1471, "finished": true, "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740880273.540, "dur": 0.169,
    "args": {
      "Python parent id": 1471, "Python id": 1472, "finished": true, "Ev Idx": 4372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740880275.470, "dur": 0.842,
    "args": {
      "Python parent id": 1469, "Python id": 1473, "finished": true, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(494): __enter__", "pid": 49625, "tid": 49625,
    "ts": 7253740880289.825, "dur": 2.605,
    "args": {
      "Python parent id": 1437, "Python id": 1474, "finished": true, "Ev Idx": 4374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_exchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740880291.269, "dur": 0.962,
    "args": {
      "Python parent id": 1474, "Python id": 1475, "finished": true, "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_synchronize>", "pid": 49625, "tid": 49625,
    "ts": 7253740880294.161, "dur": 20579.320,
    "args": {
      "Python parent id": 1437, "Python id": 1476, "finished": true, "Ev Idx": 4376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(497): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740900875.025, "dur": 3.952,
    "args": {
      "Python parent id": 1437, "Python id": 1477, "finished": true, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_maybeExchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740900877.771, "dur": 0.918,
    "args": {
      "Python parent id": 1477, "Python id": 1478, "finished": true, "Ev Idx": 4378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900887.126, "dur": 8.459,
    "args": {
      "Python parent id": 12, "Python id": 1479, "finished": true, "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900891.750, "dur": 3.595,
    "args": {
      "Python parent id": 1479, "Python id": 1480, "finished": true, "Ev Idx": 4380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900897.403, "dur": 1.128,
    "args": {
      "Python parent id": 12, "Python id": 1481, "finished": true, "Ev Idx": 4381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900897.877, "dur": 0.602,
    "args": {
      "Python parent id": 1481, "Python id": 1482, "finished": true, "Ev Idx": 4382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900899.363, "dur": 0.852,
    "args": {
      "Python parent id": 12, "Python id": 1483, "finished": true, "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900899.626, "dur": 0.539,
    "args": {
      "Python parent id": 1483, "Python id": 1484, "finished": true, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900900.840, "dur": 0.671,
    "args": {
      "Python parent id": 12, "Python id": 1485, "finished": true, "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900901.071, "dur": 0.403,
    "args": {
      "Python parent id": 1485, "Python id": 1486, "finished": true, "Ev Idx": 4386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900902.108, "dur": 1.020,
    "args": {
      "Python parent id": 12, "Python id": 1487, "finished": true, "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900902.456, "dur": 0.643,
    "args": {
      "Python parent id": 1487, "Python id": 1488, "finished": true, "Ev Idx": 4388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900903.864, "dur": 1.157,
    "args": {
      "Python parent id": 12, "Python id": 1489, "finished": true, "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900904.182, "dur": 0.784,
    "args": {
      "Python parent id": 1489, "Python id": 1490, "finished": true, "Ev Idx": 4390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900905.628, "dur": 0.836,
    "args": {
      "Python parent id": 12, "Python id": 1491, "finished": true, "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900905.847, "dur": 0.585,
    "args": {
      "Python parent id": 1491, "Python id": 1492, "finished": true, "Ev Idx": 4392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900906.970, "dur": 0.706,
    "args": {
      "Python parent id": 12, "Python id": 1493, "finished": true, "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900907.203, "dur": 0.444,
    "args": {
      "Python parent id": 1493, "Python id": 1494, "finished": true, "Ev Idx": 4394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900908.235, "dur": 0.770,
    "args": {
      "Python parent id": 12, "Python id": 1495, "finished": true, "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900908.522, "dur": 0.444,
    "args": {
      "Python parent id": 1495, "Python id": 1496, "finished": true, "Ev Idx": 4396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900909.627, "dur": 0.729,
    "args": {
      "Python parent id": 12, "Python id": 1497, "finished": true, "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900909.852, "dur": 0.474,
    "args": {
      "Python parent id": 1497, "Python id": 1498, "finished": true, "Ev Idx": 4398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900910.898, "dur": 0.687,
    "args": {
      "Python parent id": 12, "Python id": 1499, "finished": true, "Ev Idx": 4399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900911.231, "dur": 0.320,
    "args": {
      "Python parent id": 1499, "Python id": 1500, "finished": true, "Ev Idx": 4400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900912.176, "dur": 0.558,
    "args": {
      "Python parent id": 12, "Python id": 1501, "finished": true, "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900912.374, "dur": 0.320,
    "args": {
      "Python parent id": 1501, "Python id": 1502, "finished": true, "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900913.293, "dur": 0.639,
    "args": {
      "Python parent id": 12, "Python id": 1503, "finished": true, "Ev Idx": 4403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900913.573, "dur": 0.321,
    "args": {
      "Python parent id": 1503, "Python id": 1504, "finished": true, "Ev Idx": 4404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900914.444, "dur": 0.549,
    "args": {
      "Python parent id": 12, "Python id": 1505, "finished": true, "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900914.636, "dur": 0.321,
    "args": {
      "Python parent id": 1505, "Python id": 1506, "finished": true, "Ev Idx": 4406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900915.561, "dur": 0.873,
    "args": {
      "Python parent id": 12, "Python id": 1507, "finished": true, "Ev Idx": 4407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900915.919, "dur": 0.484,
    "args": {
      "Python parent id": 1507, "Python id": 1508, "finished": true, "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900916.992, "dur": 0.809,
    "args": {
      "Python parent id": 12, "Python id": 1509, "finished": true, "Ev Idx": 4409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900917.190, "dur": 0.582,
    "args": {
      "Python parent id": 1509, "Python id": 1510, "finished": true, "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900918.369, "dur": 0.888,
    "args": {
      "Python parent id": 12, "Python id": 1511, "finished": true, "Ev Idx": 4411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900918.596, "dur": 0.593,
    "args": {
      "Python parent id": 1511, "Python id": 1512, "finished": true, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900919.802, "dur": 0.701,
    "args": {
      "Python parent id": 12, "Python id": 1513, "finished": true, "Ev Idx": 4413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900920.015, "dur": 0.457,
    "args": {
      "Python parent id": 1513, "Python id": 1514, "finished": true, "Ev Idx": 4414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900921.042, "dur": 0.706,
    "args": {
      "Python parent id": 12, "Python id": 1515, "finished": true, "Ev Idx": 4415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900921.234, "dur": 0.482,
    "args": {
      "Python parent id": 1515, "Python id": 1516, "finished": true, "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900922.261, "dur": 0.818,
    "args": {
      "Python parent id": 12, "Python id": 1517, "finished": true, "Ev Idx": 4417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900922.452, "dur": 0.599,
    "args": {
      "Python parent id": 1517, "Python id": 1518, "finished": true, "Ev Idx": 4418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900923.617, "dur": 0.576,
    "args": {
      "Python parent id": 12, "Python id": 1519, "finished": true, "Ev Idx": 4419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900923.805, "dur": 0.322,
    "args": {
      "Python parent id": 1519, "Python id": 1520, "finished": true, "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900924.707, "dur": 0.552,
    "args": {
      "Python parent id": 12, "Python id": 1521, "finished": true, "Ev Idx": 4421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900924.902, "dur": 0.322,
    "args": {
      "Python parent id": 1521, "Python id": 1522, "finished": true, "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900925.778, "dur": 0.554,
    "args": {
      "Python parent id": 12, "Python id": 1523, "finished": true, "Ev Idx": 4423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900925.978, "dur": 0.322,
    "args": {
      "Python parent id": 1523, "Python id": 1524, "finished": true, "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900926.874, "dur": 0.563,
    "args": {
      "Python parent id": 12, "Python id": 1525, "finished": true, "Ev Idx": 4425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900927.065, "dur": 0.335,
    "args": {
      "Python parent id": 1525, "Python id": 1526, "finished": true, "Ev Idx": 4426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900927.978, "dur": 0.547,
    "args": {
      "Python parent id": 12, "Python id": 1527, "finished": true, "Ev Idx": 4427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900928.175, "dur": 0.319,
    "args": {
      "Python parent id": 1527, "Python id": 1528, "finished": true, "Ev Idx": 4428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900929.430, "dur": 0.548,
    "args": {
      "Python parent id": 12, "Python id": 1529, "finished": true, "Ev Idx": 4429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900929.622, "dur": 0.318,
    "args": {
      "Python parent id": 1529, "Python id": 1530, "finished": true, "Ev Idx": 4430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900930.576, "dur": 0.548,
    "args": {
      "Python parent id": 12, "Python id": 1531, "finished": true, "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900930.766, "dur": 0.317,
    "args": {
      "Python parent id": 1531, "Python id": 1532, "finished": true, "Ev Idx": 4432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900931.678, "dur": 0.548,
    "args": {
      "Python parent id": 12, "Python id": 1533, "finished": true, "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900931.868, "dur": 0.316,
    "args": {
      "Python parent id": 1533, "Python id": 1534, "finished": true, "Ev Idx": 4434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900932.754, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1535, "finished": true, "Ev Idx": 4435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900932.948, "dur": 0.317,
    "args": {
      "Python parent id": 1535, "Python id": 1536, "finished": true, "Ev Idx": 4436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900933.862, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1537, "finished": true, "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900934.056, "dur": 0.317,
    "args": {
      "Python parent id": 1537, "Python id": 1538, "finished": true, "Ev Idx": 4438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900934.946, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1539, "finished": true, "Ev Idx": 4439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900935.138, "dur": 0.319,
    "args": {
      "Python parent id": 1539, "Python id": 1540, "finished": true, "Ev Idx": 4440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900936.043, "dur": 0.574,
    "args": {
      "Python parent id": 12, "Python id": 1541, "finished": true, "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900936.236, "dur": 0.322,
    "args": {
      "Python parent id": 1541, "Python id": 1542, "finished": true, "Ev Idx": 4442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900937.199, "dur": 0.590,
    "args": {
      "Python parent id": 12, "Python id": 1543, "finished": true, "Ev Idx": 4443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900937.434, "dur": 0.318,
    "args": {
      "Python parent id": 1543, "Python id": 1544, "finished": true, "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900938.735, "dur": 0.573,
    "args": {
      "Python parent id": 12, "Python id": 1545, "finished": true, "Ev Idx": 4445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900938.926, "dur": 0.344,
    "args": {
      "Python parent id": 1545, "Python id": 1546, "finished": true, "Ev Idx": 4446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900939.837, "dur": 0.580,
    "args": {
      "Python parent id": 12, "Python id": 1547, "finished": true, "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900940.033, "dur": 0.319,
    "args": {
      "Python parent id": 1547, "Python id": 1548, "finished": true, "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900941.059, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1549, "finished": true, "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900941.256, "dur": 0.308,
    "args": {
      "Python parent id": 1549, "Python id": 1550, "finished": true, "Ev Idx": 4450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900942.117, "dur": 0.540,
    "args": {
      "Python parent id": 12, "Python id": 1551, "finished": true, "Ev Idx": 4451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900942.308, "dur": 0.312,
    "args": {
      "Python parent id": 1551, "Python id": 1552, "finished": true, "Ev Idx": 4452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900943.189, "dur": 0.847,
    "args": {
      "Python parent id": 12, "Python id": 1553, "finished": true, "Ev Idx": 4453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900943.388, "dur": 0.580,
    "args": {
      "Python parent id": 1553, "Python id": 1554, "finished": true, "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900944.605, "dur": 0.605,
    "args": {
      "Python parent id": 12, "Python id": 1555, "finished": true, "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900944.810, "dur": 0.364,
    "args": {
      "Python parent id": 1555, "Python id": 1556, "finished": true, "Ev Idx": 4456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900945.750, "dur": 0.574,
    "args": {
      "Python parent id": 12, "Python id": 1557, "finished": true, "Ev Idx": 4457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900945.963, "dur": 0.320,
    "args": {
      "Python parent id": 1557, "Python id": 1558, "finished": true, "Ev Idx": 4458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900946.878, "dur": 0.558,
    "args": {
      "Python parent id": 12, "Python id": 1559, "finished": true, "Ev Idx": 4459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900947.076, "dur": 0.319,
    "args": {
      "Python parent id": 1559, "Python id": 1560, "finished": true, "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900948.296, "dur": 0.578,
    "args": {
      "Python parent id": 12, "Python id": 1561, "finished": true, "Ev Idx": 4461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900948.496, "dur": 0.347,
    "args": {
      "Python parent id": 1561, "Python id": 1562, "finished": true, "Ev Idx": 4462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900949.402, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1563, "finished": true, "Ev Idx": 4463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900949.596, "dur": 0.318,
    "args": {
      "Python parent id": 1563, "Python id": 1564, "finished": true, "Ev Idx": 4464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900950.587, "dur": 0.547,
    "args": {
      "Python parent id": 12, "Python id": 1565, "finished": true, "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900950.782, "dur": 0.320,
    "args": {
      "Python parent id": 1565, "Python id": 1566, "finished": true, "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900951.663, "dur": 0.551,
    "args": {
      "Python parent id": 12, "Python id": 1567, "finished": true, "Ev Idx": 4467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900951.860, "dur": 0.319,
    "args": {
      "Python parent id": 1567, "Python id": 1568, "finished": true, "Ev Idx": 4468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900952.736, "dur": 0.543,
    "args": {
      "Python parent id": 12, "Python id": 1569, "finished": true, "Ev Idx": 4469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900952.933, "dur": 0.313,
    "args": {
      "Python parent id": 1569, "Python id": 1570, "finished": true, "Ev Idx": 4470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900953.875, "dur": 0.595,
    "args": {
      "Python parent id": 12, "Python id": 1571, "finished": true, "Ev Idx": 4471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900954.103, "dur": 0.315,
    "args": {
      "Python parent id": 1571, "Python id": 1572, "finished": true, "Ev Idx": 4472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900955.013, "dur": 0.543,
    "args": {
      "Python parent id": 12, "Python id": 1573, "finished": true, "Ev Idx": 4473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900955.208, "dur": 0.314,
    "args": {
      "Python parent id": 1573, "Python id": 1574, "finished": true, "Ev Idx": 4474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900956.081, "dur": 0.552,
    "args": {
      "Python parent id": 12, "Python id": 1575, "finished": true, "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900956.286, "dur": 0.315,
    "args": {
      "Python parent id": 1575, "Python id": 1576, "finished": true, "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900957.162, "dur": 0.546,
    "args": {
      "Python parent id": 12, "Python id": 1577, "finished": true, "Ev Idx": 4477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900957.362, "dur": 0.313,
    "args": {
      "Python parent id": 1577, "Python id": 1578, "finished": true, "Ev Idx": 4478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900958.279, "dur": 0.676,
    "args": {
      "Python parent id": 12, "Python id": 1579, "finished": true, "Ev Idx": 4479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900958.473, "dur": 0.443,
    "args": {
      "Python parent id": 1579, "Python id": 1580, "finished": true, "Ev Idx": 4480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900959.603, "dur": 0.802,
    "args": {
      "Python parent id": 12, "Python id": 1581, "finished": true, "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900959.798, "dur": 0.576,
    "args": {
      "Python parent id": 1581, "Python id": 1582, "finished": true, "Ev Idx": 4482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900960.954, "dur": 1.114,
    "args": {
      "Python parent id": 12, "Python id": 1583, "finished": true, "Ev Idx": 4483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900961.152, "dur": 0.886,
    "args": {
      "Python parent id": 1583, "Python id": 1584, "finished": true, "Ev Idx": 4484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900963.109, "dur": 0.738,
    "args": {
      "Python parent id": 12, "Python id": 1585, "finished": true, "Ev Idx": 4485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900963.309, "dur": 0.509,
    "args": {
      "Python parent id": 1585, "Python id": 1586, "finished": true, "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900964.470, "dur": 0.733,
    "args": {
      "Python parent id": 12, "Python id": 1587, "finished": true, "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900964.718, "dur": 0.454,
    "args": {
      "Python parent id": 1587, "Python id": 1588, "finished": true, "Ev Idx": 4488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900965.723, "dur": 0.774,
    "args": {
      "Python parent id": 12, "Python id": 1589, "finished": true, "Ev Idx": 4489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900965.919, "dur": 0.548,
    "args": {
      "Python parent id": 1589, "Python id": 1590, "finished": true, "Ev Idx": 4490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900967.042, "dur": 0.700,
    "args": {
      "Python parent id": 12, "Python id": 1591, "finished": true, "Ev Idx": 4491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900967.240, "dur": 0.473,
    "args": {
      "Python parent id": 1591, "Python id": 1592, "finished": true, "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900968.302, "dur": 0.558,
    "args": {
      "Python parent id": 12, "Python id": 1593, "finished": true, "Ev Idx": 4493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900968.500, "dur": 0.321,
    "args": {
      "Python parent id": 1593, "Python id": 1594, "finished": true, "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900969.490, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1595, "finished": true, "Ev Idx": 4495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900969.687, "dur": 0.317,
    "args": {
      "Python parent id": 1595, "Python id": 1596, "finished": true, "Ev Idx": 4496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900970.574, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1597, "finished": true, "Ev Idx": 4497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900970.769, "dur": 0.309,
    "args": {
      "Python parent id": 1597, "Python id": 1598, "finished": true, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900971.651, "dur": 0.611,
    "args": {
      "Python parent id": 12, "Python id": 1599, "finished": true, "Ev Idx": 4499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900971.880, "dur": 0.315,
    "args": {
      "Python parent id": 1599, "Python id": 1600, "finished": true, "Ev Idx": 4500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900972.787, "dur": 0.549,
    "args": {
      "Python parent id": 12, "Python id": 1601, "finished": true, "Ev Idx": 4501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900972.986, "dur": 0.317,
    "args": {
      "Python parent id": 1601, "Python id": 1602, "finished": true, "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900973.890, "dur": 0.546,
    "args": {
      "Python parent id": 12, "Python id": 1603, "finished": true, "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900974.082, "dur": 0.317,
    "args": {
      "Python parent id": 1603, "Python id": 1604, "finished": true, "Ev Idx": 4504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900974.947, "dur": 0.552,
    "args": {
      "Python parent id": 12, "Python id": 1605, "finished": true, "Ev Idx": 4505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900975.141, "dur": 0.316,
    "args": {
      "Python parent id": 1605, "Python id": 1606, "finished": true, "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900976.066, "dur": 0.579,
    "args": {
      "Python parent id": 12, "Python id": 1607, "finished": true, "Ev Idx": 4507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900976.294, "dur": 0.312,
    "args": {
      "Python parent id": 1607, "Python id": 1608, "finished": true, "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900977.869, "dur": 0.543,
    "args": {
      "Python parent id": 12, "Python id": 1609, "finished": true, "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900978.065, "dur": 0.312,
    "args": {
      "Python parent id": 1609, "Python id": 1610, "finished": true, "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900978.942, "dur": 0.597,
    "args": {
      "Python parent id": 12, "Python id": 1611, "finished": true, "Ev Idx": 4511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900979.136, "dur": 0.366,
    "args": {
      "Python parent id": 1611, "Python id": 1612, "finished": true, "Ev Idx": 4512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900980.098, "dur": 0.563,
    "args": {
      "Python parent id": 12, "Python id": 1613, "finished": true, "Ev Idx": 4513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900980.309, "dur": 0.318,
    "args": {
      "Python parent id": 1613, "Python id": 1614, "finished": true, "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900981.388, "dur": 0.567,
    "args": {
      "Python parent id": 12, "Python id": 1615, "finished": true, "Ev Idx": 4515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900981.584, "dur": 0.313,
    "args": {
      "Python parent id": 1615, "Python id": 1616, "finished": true, "Ev Idx": 4516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900982.509, "dur": 0.570,
    "args": {
      "Python parent id": 12, "Python id": 1617, "finished": true, "Ev Idx": 4517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900982.701, "dur": 0.341,
    "args": {
      "Python parent id": 1617, "Python id": 1618, "finished": true, "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900983.620, "dur": 0.553,
    "args": {
      "Python parent id": 12, "Python id": 1619, "finished": true, "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900983.820, "dur": 0.316,
    "args": {
      "Python parent id": 1619, "Python id": 1620, "finished": true, "Ev Idx": 4520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900984.853, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1621, "finished": true, "Ev Idx": 4521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900985.047, "dur": 0.318,
    "args": {
      "Python parent id": 1621, "Python id": 1622, "finished": true, "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900985.957, "dur": 0.551,
    "args": {
      "Python parent id": 12, "Python id": 1623, "finished": true, "Ev Idx": 4523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900986.152, "dur": 0.317,
    "args": {
      "Python parent id": 1623, "Python id": 1624, "finished": true, "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900987.054, "dur": 0.594,
    "args": {
      "Python parent id": 12, "Python id": 1625, "finished": true, "Ev Idx": 4525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900987.248, "dur": 0.342,
    "args": {
      "Python parent id": 1625, "Python id": 1626, "finished": true, "Ev Idx": 4526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900988.177, "dur": 0.559,
    "args": {
      "Python parent id": 12, "Python id": 1627, "finished": true, "Ev Idx": 4527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900988.376, "dur": 0.323,
    "args": {
      "Python parent id": 1627, "Python id": 1628, "finished": true, "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900989.248, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1629, "finished": true, "Ev Idx": 4529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900989.443, "dur": 0.310,
    "args": {
      "Python parent id": 1629, "Python id": 1630, "finished": true, "Ev Idx": 4530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740900990.320, "dur": 0.583,
    "args": {
      "Python parent id": 12, "Python id": 1631, "finished": true, "Ev Idx": 4531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740900990.511, "dur": 0.335,
    "args": {
      "Python parent id": 1631, "Python id": 1632, "finished": true, "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901000.306, "dur": 1.387,
    "args": {
      "Python parent id": 12, "Python id": 1633, "finished": true, "Ev Idx": 4533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901000.866, "dur": 0.770,
    "args": {
      "Python parent id": 1633, "Python id": 1634, "finished": true, "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901002.311, "dur": 0.638,
    "args": {
      "Python parent id": 12, "Python id": 1635, "finished": true, "Ev Idx": 4535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901002.556, "dur": 0.351,
    "args": {
      "Python parent id": 1635, "Python id": 1636, "finished": true, "Ev Idx": 4536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901003.483, "dur": 0.593,
    "args": {
      "Python parent id": 12, "Python id": 1637, "finished": true, "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901003.683, "dur": 0.359,
    "args": {
      "Python parent id": 1637, "Python id": 1638, "finished": true, "Ev Idx": 4538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901004.633, "dur": 0.587,
    "args": {
      "Python parent id": 12, "Python id": 1639, "finished": true, "Ev Idx": 4539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901004.835, "dur": 0.325,
    "args": {
      "Python parent id": 1639, "Python id": 1640, "finished": true, "Ev Idx": 4540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901005.739, "dur": 0.605,
    "args": {
      "Python parent id": 12, "Python id": 1641, "finished": true, "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901005.984, "dur": 0.320,
    "args": {
      "Python parent id": 1641, "Python id": 1642, "finished": true, "Ev Idx": 4542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901006.904, "dur": 0.546,
    "args": {
      "Python parent id": 12, "Python id": 1643, "finished": true, "Ev Idx": 4543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901007.100, "dur": 0.312,
    "args": {
      "Python parent id": 1643, "Python id": 1644, "finished": true, "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901007.980, "dur": 0.551,
    "args": {
      "Python parent id": 12, "Python id": 1645, "finished": true, "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901008.187, "dur": 0.309,
    "args": {
      "Python parent id": 1645, "Python id": 1646, "finished": true, "Ev Idx": 4546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901009.077, "dur": 0.545,
    "args": {
      "Python parent id": 12, "Python id": 1647, "finished": true, "Ev Idx": 4547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901009.277, "dur": 0.311,
    "args": {
      "Python parent id": 1647, "Python id": 1648, "finished": true, "Ev Idx": 4548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901020.746, "dur": 1.283,
    "args": {
      "Python parent id": 12, "Python id": 1649, "finished": true, "Ev Idx": 4549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901021.273, "dur": 0.719,
    "args": {
      "Python parent id": 1649, "Python id": 1650, "finished": true, "Ev Idx": 4550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901022.672, "dur": 0.627,
    "args": {
      "Python parent id": 12, "Python id": 1651, "finished": true, "Ev Idx": 4551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901022.908, "dur": 0.349,
    "args": {
      "Python parent id": 1651, "Python id": 1652, "finished": true, "Ev Idx": 4552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901023.833, "dur": 0.860,
    "args": {
      "Python parent id": 12, "Python id": 1653, "finished": true, "Ev Idx": 4553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901024.038, "dur": 0.624,
    "args": {
      "Python parent id": 1653, "Python id": 1654, "finished": true, "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901025.241, "dur": 0.866,
    "args": {
      "Python parent id": 12, "Python id": 1655, "finished": true, "Ev Idx": 4555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901025.453, "dur": 0.624,
    "args": {
      "Python parent id": 1655, "Python id": 1656, "finished": true, "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901026.674, "dur": 0.699,
    "args": {
      "Python parent id": 12, "Python id": 1657, "finished": true, "Ev Idx": 4557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901026.879, "dur": 0.461,
    "args": {
      "Python parent id": 1657, "Python id": 1658, "finished": true, "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901027.955, "dur": 0.708,
    "args": {
      "Python parent id": 12, "Python id": 1659, "finished": true, "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901028.159, "dur": 0.471,
    "args": {
      "Python parent id": 1659, "Python id": 1660, "finished": true, "Ev Idx": 4560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901029.226, "dur": 0.782,
    "args": {
      "Python parent id": 12, "Python id": 1661, "finished": true, "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901029.427, "dur": 0.553,
    "args": {
      "Python parent id": 1661, "Python id": 1662, "finished": true, "Ev Idx": 4562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901030.533, "dur": 0.769,
    "args": {
      "Python parent id": 12, "Python id": 1663, "finished": true, "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901030.733, "dur": 0.540,
    "args": {
      "Python parent id": 1663, "Python id": 1664, "finished": true, "Ev Idx": 4564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901037.027, "dur": 1.293,
    "args": {
      "Python parent id": 12, "Python id": 1665, "finished": true, "Ev Idx": 4565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901037.547, "dur": 0.714,
    "args": {
      "Python parent id": 1665, "Python id": 1666, "finished": true, "Ev Idx": 4566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901038.927, "dur": 0.673,
    "args": {
      "Python parent id": 12, "Python id": 1667, "finished": true, "Ev Idx": 4567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901039.212, "dur": 0.345,
    "args": {
      "Python parent id": 1667, "Python id": 1668, "finished": true, "Ev Idx": 4568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901040.128, "dur": 0.688,
    "args": {
      "Python parent id": 12, "Python id": 1669, "finished": true, "Ev Idx": 4569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901040.418, "dur": 0.362,
    "args": {
      "Python parent id": 1669, "Python id": 1670, "finished": true, "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901041.378, "dur": 0.647,
    "args": {
      "Python parent id": 12, "Python id": 1671, "finished": true, "Ev Idx": 4571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901041.645, "dur": 0.346,
    "args": {
      "Python parent id": 1671, "Python id": 1672, "finished": true, "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901042.569, "dur": 0.591,
    "args": {
      "Python parent id": 12, "Python id": 1673, "finished": true, "Ev Idx": 4573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901042.799, "dur": 0.326,
    "args": {
      "Python parent id": 1673, "Python id": 1674, "finished": true, "Ev Idx": 4574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901043.681, "dur": 0.625,
    "args": {
      "Python parent id": 12, "Python id": 1675, "finished": true, "Ev Idx": 4575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901043.923, "dur": 0.342,
    "args": {
      "Python parent id": 1675, "Python id": 1676, "finished": true, "Ev Idx": 4576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901044.871, "dur": 0.611,
    "args": {
      "Python parent id": 12, "Python id": 1677, "finished": true, "Ev Idx": 4577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901045.103, "dur": 0.345,
    "args": {
      "Python parent id": 1677, "Python id": 1678, "finished": true, "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901047.278, "dur": 0.710,
    "args": {
      "Python parent id": 12, "Python id": 1679, "finished": true, "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901047.592, "dur": 0.362,
    "args": {
      "Python parent id": 1679, "Python id": 1680, "finished": true, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901048.959, "dur": 0.641,
    "args": {
      "Python parent id": 12, "Python id": 1681, "finished": true, "Ev Idx": 4581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901049.228, "dur": 0.334,
    "args": {
      "Python parent id": 1681, "Python id": 1682, "finished": true, "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901050.294, "dur": 0.615,
    "args": {
      "Python parent id": 12, "Python id": 1683, "finished": true, "Ev Idx": 4583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901050.523, "dur": 0.335,
    "args": {
      "Python parent id": 1683, "Python id": 1684, "finished": true, "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901051.517, "dur": 0.573,
    "args": {
      "Python parent id": 12, "Python id": 1685, "finished": true, "Ev Idx": 4585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901051.743, "dur": 0.313,
    "args": {
      "Python parent id": 1685, "Python id": 1686, "finished": true, "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901052.713, "dur": 0.580,
    "args": {
      "Python parent id": 12, "Python id": 1687, "finished": true, "Ev Idx": 4587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901052.934, "dur": 0.326,
    "args": {
      "Python parent id": 1687, "Python id": 1688, "finished": true, "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901053.981, "dur": 0.686,
    "args": {
      "Python parent id": 12, "Python id": 1689, "finished": true, "Ev Idx": 4589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901054.242, "dur": 0.393,
    "args": {
      "Python parent id": 1689, "Python id": 1690, "finished": true, "Ev Idx": 4590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901055.320, "dur": 0.688,
    "args": {
      "Python parent id": 12, "Python id": 1691, "finished": true, "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901055.616, "dur": 0.338,
    "args": {
      "Python parent id": 1691, "Python id": 1692, "finished": true, "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901056.550, "dur": 0.635,
    "args": {
      "Python parent id": 12, "Python id": 1693, "finished": true, "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901056.790, "dur": 0.359,
    "args": {
      "Python parent id": 1693, "Python id": 1694, "finished": true, "Ev Idx": 4594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901057.728, "dur": 0.596,
    "args": {
      "Python parent id": 12, "Python id": 1695, "finished": true, "Ev Idx": 4595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901057.959, "dur": 0.334,
    "args": {
      "Python parent id": 1695, "Python id": 1696, "finished": true, "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901058.927, "dur": 0.652,
    "args": {
      "Python parent id": 12, "Python id": 1697, "finished": true, "Ev Idx": 4597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901059.173, "dur": 0.367,
    "args": {
      "Python parent id": 1697, "Python id": 1698, "finished": true, "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901060.145, "dur": 0.593,
    "args": {
      "Python parent id": 12, "Python id": 1699, "finished": true, "Ev Idx": 4599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901060.390, "dur": 0.310,
    "args": {
      "Python parent id": 1699, "Python id": 1700, "finished": true, "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901061.335, "dur": 164.143,
    "args": {
      "Python parent id": 12, "Python id": 1701, "finished": true, "Ev Idx": 4601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901061.644, "dur": 0.362,
    "args": {
      "Python parent id": 1701, "Python id": 1702, "finished": true, "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901226.632, "dur": 1.236,
    "args": {
      "Python parent id": 12, "Python id": 1703, "finished": true, "Ev Idx": 4603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901227.153, "dur": 0.605,
    "args": {
      "Python parent id": 1703, "Python id": 1704, "finished": true, "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901228.656, "dur": 0.754,
    "args": {
      "Python parent id": 12, "Python id": 1705, "finished": true, "Ev Idx": 4605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901228.939, "dur": 0.404,
    "args": {
      "Python parent id": 1705, "Python id": 1706, "finished": true, "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901230.053, "dur": 0.592,
    "args": {
      "Python parent id": 12, "Python id": 1707, "finished": true, "Ev Idx": 4607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901230.271, "dur": 0.323,
    "args": {
      "Python parent id": 1707, "Python id": 1708, "finished": true, "Ev Idx": 4608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901231.232, "dur": 0.670,
    "args": {
      "Python parent id": 12, "Python id": 1709, "finished": true, "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901231.470, "dur": 0.405,
    "args": {
      "Python parent id": 1709, "Python id": 1710, "finished": true, "Ev Idx": 4610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901232.459, "dur": 0.604,
    "args": {
      "Python parent id": 12, "Python id": 1711, "finished": true, "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901232.693, "dur": 0.343,
    "args": {
      "Python parent id": 1711, "Python id": 1712, "finished": true, "Ev Idx": 4612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901233.813, "dur": 0.602,
    "args": {
      "Python parent id": 12, "Python id": 1713, "finished": true, "Ev Idx": 4613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901234.029, "dur": 0.336,
    "args": {
      "Python parent id": 1713, "Python id": 1714, "finished": true, "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901235.026, "dur": 0.749,
    "args": {
      "Python parent id": 12, "Python id": 1715, "finished": true, "Ev Idx": 4615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901235.275, "dur": 0.470,
    "args": {
      "Python parent id": 1715, "Python id": 1716, "finished": true, "Ev Idx": 4616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901236.381, "dur": 0.573,
    "args": {
      "Python parent id": 12, "Python id": 1717, "finished": true, "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901236.598, "dur": 0.320,
    "args": {
      "Python parent id": 1717, "Python id": 1718, "finished": true, "Ev Idx": 4618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901237.529, "dur": 0.570,
    "args": {
      "Python parent id": 12, "Python id": 1719, "finished": true, "Ev Idx": 4619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901237.745, "dur": 0.318,
    "args": {
      "Python parent id": 1719, "Python id": 1720, "finished": true, "Ev Idx": 4620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901238.672, "dur": 0.568,
    "args": {
      "Python parent id": 12, "Python id": 1721, "finished": true, "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901238.886, "dur": 0.317,
    "args": {
      "Python parent id": 1721, "Python id": 1722, "finished": true, "Ev Idx": 4622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901239.782, "dur": 0.622,
    "args": {
      "Python parent id": 12, "Python id": 1723, "finished": true, "Ev Idx": 4623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901240.000, "dur": 0.366,
    "args": {
      "Python parent id": 1723, "Python id": 1724, "finished": true, "Ev Idx": 4624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901241.042, "dur": 0.695,
    "args": {
      "Python parent id": 12, "Python id": 1725, "finished": true, "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901241.267, "dur": 0.437,
    "args": {
      "Python parent id": 1725, "Python id": 1726, "finished": true, "Ev Idx": 4626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901242.304, "dur": 0.579,
    "args": {
      "Python parent id": 12, "Python id": 1727, "finished": true, "Ev Idx": 4627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901242.521, "dur": 0.326,
    "args": {
      "Python parent id": 1727, "Python id": 1728, "finished": true, "Ev Idx": 4628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901243.512, "dur": 0.661,
    "args": {
      "Python parent id": 12, "Python id": 1729, "finished": true, "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901243.807, "dur": 0.327,
    "args": {
      "Python parent id": 1729, "Python id": 1730, "finished": true, "Ev Idx": 4630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901244.792, "dur": 0.717,
    "args": {
      "Python parent id": 12, "Python id": 1731, "finished": true, "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901245.010, "dur": 0.447,
    "args": {
      "Python parent id": 1731, "Python id": 1732, "finished": true, "Ev Idx": 4632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901246.077, "dur": 0.680,
    "args": {
      "Python parent id": 12, "Python id": 1733, "finished": true, "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901246.292, "dur": 0.434,
    "args": {
      "Python parent id": 1733, "Python id": 1734, "finished": true, "Ev Idx": 4634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901247.365, "dur": 0.772,
    "args": {
      "Python parent id": 12, "Python id": 1735, "finished": true, "Ev Idx": 4635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901247.586, "dur": 0.524,
    "args": {
      "Python parent id": 1735, "Python id": 1736, "finished": true, "Ev Idx": 4636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901248.756, "dur": 0.632,
    "args": {
      "Python parent id": 12, "Python id": 1737, "finished": true, "Ev Idx": 4637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901248.976, "dur": 0.385,
    "args": {
      "Python parent id": 1737, "Python id": 1738, "finished": true, "Ev Idx": 4638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901249.959, "dur": 0.584,
    "args": {
      "Python parent id": 12, "Python id": 1739, "finished": true, "Ev Idx": 4639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901250.179, "dur": 0.333,
    "args": {
      "Python parent id": 1739, "Python id": 1740, "finished": true, "Ev Idx": 4640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901251.183, "dur": 0.665,
    "args": {
      "Python parent id": 12, "Python id": 1741, "finished": true, "Ev Idx": 4641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901251.399, "dur": 0.416,
    "args": {
      "Python parent id": 1741, "Python id": 1742, "finished": true, "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901252.410, "dur": 0.592,
    "args": {
      "Python parent id": 12, "Python id": 1743, "finished": true, "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901252.633, "dur": 0.309,
    "args": {
      "Python parent id": 1743, "Python id": 1744, "finished": true, "Ev Idx": 4644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901253.617, "dur": 0.563,
    "args": {
      "Python parent id": 12, "Python id": 1745, "finished": true, "Ev Idx": 4645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901253.836, "dur": 0.307,
    "args": {
      "Python parent id": 1745, "Python id": 1746, "finished": true, "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901254.751, "dur": 0.574,
    "args": {
      "Python parent id": 12, "Python id": 1747, "finished": true, "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901254.966, "dur": 0.319,
    "args": {
      "Python parent id": 1747, "Python id": 1748, "finished": true, "Ev Idx": 4648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901255.928, "dur": 0.563,
    "args": {
      "Python parent id": 12, "Python id": 1749, "finished": true, "Ev Idx": 4649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901256.147, "dur": 0.309,
    "args": {
      "Python parent id": 1749, "Python id": 1750, "finished": true, "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901257.087, "dur": 0.592,
    "args": {
      "Python parent id": 12, "Python id": 1751, "finished": true, "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901257.303, "dur": 0.323,
    "args": {
      "Python parent id": 1751, "Python id": 1752, "finished": true, "Ev Idx": 4652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901258.241, "dur": 0.601,
    "args": {
      "Python parent id": 12, "Python id": 1753, "finished": true, "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901258.494, "dur": 0.314,
    "args": {
      "Python parent id": 1753, "Python id": 1754, "finished": true, "Ev Idx": 4654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901259.412, "dur": 0.572,
    "args": {
      "Python parent id": 12, "Python id": 1755, "finished": true, "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901259.632, "dur": 0.311,
    "args": {
      "Python parent id": 1755, "Python id": 1756, "finished": true, "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901260.615, "dur": 0.576,
    "args": {
      "Python parent id": 12, "Python id": 1757, "finished": true, "Ev Idx": 4657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901260.836, "dur": 0.313,
    "args": {
      "Python parent id": 1757, "Python id": 1758, "finished": true, "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901261.758, "dur": 0.570,
    "args": {
      "Python parent id": 12, "Python id": 1759, "finished": true, "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901261.976, "dur": 0.311,
    "args": {
      "Python parent id": 1759, "Python id": 1760, "finished": true, "Ev Idx": 4660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901262.936, "dur": 0.584,
    "args": {
      "Python parent id": 12, "Python id": 1761, "finished": true, "Ev Idx": 4661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901263.157, "dur": 0.303,
    "args": {
      "Python parent id": 1761, "Python id": 1762, "finished": true, "Ev Idx": 4662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901264.092, "dur": 0.594,
    "args": {
      "Python parent id": 12, "Python id": 1763, "finished": true, "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901264.336, "dur": 0.312,
    "args": {
      "Python parent id": 1763, "Python id": 1764, "finished": true, "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901265.274, "dur": 0.573,
    "args": {
      "Python parent id": 12, "Python id": 1765, "finished": true, "Ev Idx": 4665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901265.496, "dur": 0.310,
    "args": {
      "Python parent id": 1765, "Python id": 1766, "finished": true, "Ev Idx": 4666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901266.416, "dur": 0.583,
    "args": {
      "Python parent id": 12, "Python id": 1767, "finished": true, "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901266.647, "dur": 0.315,
    "args": {
      "Python parent id": 1767, "Python id": 1768, "finished": true, "Ev Idx": 4668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901267.578, "dur": 0.565,
    "args": {
      "Python parent id": 12, "Python id": 1769, "finished": true, "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901267.792, "dur": 0.317,
    "args": {
      "Python parent id": 1769, "Python id": 1770, "finished": true, "Ev Idx": 4670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901268.703, "dur": 0.571,
    "args": {
      "Python parent id": 12, "Python id": 1771, "finished": true, "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901268.922, "dur": 0.311,
    "args": {
      "Python parent id": 1771, "Python id": 1772, "finished": true, "Ev Idx": 4672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901269.853, "dur": 0.568,
    "args": {
      "Python parent id": 12, "Python id": 1773, "finished": true, "Ev Idx": 4673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901270.072, "dur": 0.314,
    "args": {
      "Python parent id": 1773, "Python id": 1774, "finished": true, "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901270.968, "dur": 0.589,
    "args": {
      "Python parent id": 12, "Python id": 1775, "finished": true, "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901271.186, "dur": 0.311,
    "args": {
      "Python parent id": 1775, "Python id": 1776, "finished": true, "Ev Idx": 4676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901272.136, "dur": 0.568,
    "args": {
      "Python parent id": 12, "Python id": 1777, "finished": true, "Ev Idx": 4677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901272.351, "dur": 0.314,
    "args": {
      "Python parent id": 1777, "Python id": 1778, "finished": true, "Ev Idx": 4678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901273.298, "dur": 0.559,
    "args": {
      "Python parent id": 12, "Python id": 1779, "finished": true, "Ev Idx": 4679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901273.511, "dur": 0.307,
    "args": {
      "Python parent id": 1779, "Python id": 1780, "finished": true, "Ev Idx": 4680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901274.481, "dur": 0.560,
    "args": {
      "Python parent id": 12, "Python id": 1781, "finished": true, "Ev Idx": 4681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901274.694, "dur": 0.310,
    "args": {
      "Python parent id": 1781, "Python id": 1782, "finished": true, "Ev Idx": 4682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901275.610, "dur": 0.564,
    "args": {
      "Python parent id": 12, "Python id": 1783, "finished": true, "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901275.827, "dur": 0.307,
    "args": {
      "Python parent id": 1783, "Python id": 1784, "finished": true, "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901276.848, "dur": 0.641,
    "args": {
      "Python parent id": 12, "Python id": 1785, "finished": true, "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901277.063, "dur": 0.385,
    "args": {
      "Python parent id": 1785, "Python id": 1786, "finished": true, "Ev Idx": 4686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901278.047, "dur": 0.558,
    "args": {
      "Python parent id": 12, "Python id": 1787, "finished": true, "Ev Idx": 4687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901278.262, "dur": 0.308,
    "args": {
      "Python parent id": 1787, "Python id": 1788, "finished": true, "Ev Idx": 4688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901279.164, "dur": 0.577,
    "args": {
      "Python parent id": 12, "Python id": 1789, "finished": true, "Ev Idx": 4689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901279.377, "dur": 0.308,
    "args": {
      "Python parent id": 1789, "Python id": 1790, "finished": true, "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901280.314, "dur": 0.565,
    "args": {
      "Python parent id": 12, "Python id": 1791, "finished": true, "Ev Idx": 4691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901280.536, "dur": 0.308,
    "args": {
      "Python parent id": 1791, "Python id": 1792, "finished": true, "Ev Idx": 4692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901281.498, "dur": 0.569,
    "args": {
      "Python parent id": 12, "Python id": 1793, "finished": true, "Ev Idx": 4693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901281.714, "dur": 0.314,
    "args": {
      "Python parent id": 1793, "Python id": 1794, "finished": true, "Ev Idx": 4694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901282.647, "dur": 0.580,
    "args": {
      "Python parent id": 12, "Python id": 1795, "finished": true, "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901282.864, "dur": 0.330,
    "args": {
      "Python parent id": 1795, "Python id": 1796, "finished": true, "Ev Idx": 4696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901283.859, "dur": 4.978,
    "args": {
      "Python parent id": 12, "Python id": 1797, "finished": true, "Ev Idx": 4697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901284.055, "dur": 4.700,
    "args": {
      "Python parent id": 1797, "Python id": 1798, "finished": true, "Ev Idx": 4698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901289.511, "dur": 0.700,
    "args": {
      "Python parent id": 12, "Python id": 1799, "finished": true, "Ev Idx": 4699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901289.803, "dur": 0.375,
    "args": {
      "Python parent id": 1799, "Python id": 1800, "finished": true, "Ev Idx": 4700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901290.825, "dur": 0.592,
    "args": {
      "Python parent id": 12, "Python id": 1801, "finished": true, "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901291.028, "dur": 0.352,
    "args": {
      "Python parent id": 1801, "Python id": 1802, "finished": true, "Ev Idx": 4702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901292.024, "dur": 0.571,
    "args": {
      "Python parent id": 12, "Python id": 1803, "finished": true, "Ev Idx": 4703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901292.228, "dur": 0.328,
    "args": {
      "Python parent id": 1803, "Python id": 1804, "finished": true, "Ev Idx": 4704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901293.183, "dur": 0.559,
    "args": {
      "Python parent id": 12, "Python id": 1805, "finished": true, "Ev Idx": 4705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901293.384, "dur": 0.321,
    "args": {
      "Python parent id": 1805, "Python id": 1806, "finished": true, "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901294.310, "dur": 0.548,
    "args": {
      "Python parent id": 12, "Python id": 1807, "finished": true, "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901294.506, "dur": 0.316,
    "args": {
      "Python parent id": 1807, "Python id": 1808, "finished": true, "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901295.519, "dur": 0.577,
    "args": {
      "Python parent id": 12, "Python id": 1809, "finished": true, "Ev Idx": 4709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901295.743, "dur": 0.314,
    "args": {
      "Python parent id": 1809, "Python id": 1810, "finished": true, "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901296.668, "dur": 0.608,
    "args": {
      "Python parent id": 12, "Python id": 1811, "finished": true, "Ev Idx": 4711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901296.883, "dur": 0.338,
    "args": {
      "Python parent id": 1811, "Python id": 1812, "finished": true, "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901297.878, "dur": 0.563,
    "args": {
      "Python parent id": 12, "Python id": 1813, "finished": true, "Ev Idx": 4713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901298.073, "dur": 0.332,
    "args": {
      "Python parent id": 1813, "Python id": 1814, "finished": true, "Ev Idx": 4714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901299.020, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1815, "finished": true, "Ev Idx": 4715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901299.217, "dur": 0.313,
    "args": {
      "Python parent id": 1815, "Python id": 1816, "finished": true, "Ev Idx": 4716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901300.120, "dur": 0.562,
    "args": {
      "Python parent id": 12, "Python id": 1817, "finished": true, "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901300.318, "dur": 0.310,
    "args": {
      "Python parent id": 1817, "Python id": 1818, "finished": true, "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901301.248, "dur": 0.537,
    "args": {
      "Python parent id": 12, "Python id": 1819, "finished": true, "Ev Idx": 4719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901301.443, "dur": 0.307,
    "args": {
      "Python parent id": 1819, "Python id": 1820, "finished": true, "Ev Idx": 4720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901302.362, "dur": 0.532,
    "args": {
      "Python parent id": 12, "Python id": 1821, "finished": true, "Ev Idx": 4721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901302.557, "dur": 0.306,
    "args": {
      "Python parent id": 1821, "Python id": 1822, "finished": true, "Ev Idx": 4722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901303.477, "dur": 0.533,
    "args": {
      "Python parent id": 12, "Python id": 1823, "finished": true, "Ev Idx": 4723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901303.671, "dur": 0.303,
    "args": {
      "Python parent id": 1823, "Python id": 1824, "finished": true, "Ev Idx": 4724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901304.583, "dur": 0.527,
    "args": {
      "Python parent id": 12, "Python id": 1825, "finished": true, "Ev Idx": 4725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901304.776, "dur": 0.300,
    "args": {
      "Python parent id": 1825, "Python id": 1826, "finished": true, "Ev Idx": 4726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901305.714, "dur": 0.544,
    "args": {
      "Python parent id": 12, "Python id": 1827, "finished": true, "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901305.908, "dur": 0.319,
    "args": {
      "Python parent id": 1827, "Python id": 1828, "finished": true, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901306.828, "dur": 0.532,
    "args": {
      "Python parent id": 12, "Python id": 1829, "finished": true, "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901307.022, "dur": 0.301,
    "args": {
      "Python parent id": 1829, "Python id": 1830, "finished": true, "Ev Idx": 4730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901307.937, "dur": 0.541,
    "args": {
      "Python parent id": 12, "Python id": 1831, "finished": true, "Ev Idx": 4731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901308.132, "dur": 0.312,
    "args": {
      "Python parent id": 1831, "Python id": 1832, "finished": true, "Ev Idx": 4732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901317.997, "dur": 0.683,
    "args": {
      "Python parent id": 12, "Python id": 1833, "finished": true, "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901318.270, "dur": 0.367,
    "args": {
      "Python parent id": 1833, "Python id": 1834, "finished": true, "Ev Idx": 4734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901319.280, "dur": 0.600,
    "args": {
      "Python parent id": 12, "Python id": 1835, "finished": true, "Ev Idx": 4735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901319.485, "dur": 0.342,
    "args": {
      "Python parent id": 1835, "Python id": 1836, "finished": true, "Ev Idx": 4736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901320.434, "dur": 0.574,
    "args": {
      "Python parent id": 12, "Python id": 1837, "finished": true, "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901320.657, "dur": 0.319,
    "args": {
      "Python parent id": 1837, "Python id": 1838, "finished": true, "Ev Idx": 4738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901321.588, "dur": 0.545,
    "args": {
      "Python parent id": 12, "Python id": 1839, "finished": true, "Ev Idx": 4739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901321.784, "dur": 0.319,
    "args": {
      "Python parent id": 1839, "Python id": 1840, "finished": true, "Ev Idx": 4740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901322.708, "dur": 0.564,
    "args": {
      "Python parent id": 12, "Python id": 1841, "finished": true, "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901322.901, "dur": 0.318,
    "args": {
      "Python parent id": 1841, "Python id": 1842, "finished": true, "Ev Idx": 4742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901323.846, "dur": 0.541,
    "args": {
      "Python parent id": 12, "Python id": 1843, "finished": true, "Ev Idx": 4743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901324.038, "dur": 0.317,
    "args": {
      "Python parent id": 1843, "Python id": 1844, "finished": true, "Ev Idx": 4744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901324.963, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1845, "finished": true, "Ev Idx": 4745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901325.161, "dur": 0.319,
    "args": {
      "Python parent id": 1845, "Python id": 1846, "finished": true, "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901326.083, "dur": 0.566,
    "args": {
      "Python parent id": 12, "Python id": 1847, "finished": true, "Ev Idx": 4747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901326.293, "dur": 0.324,
    "args": {
      "Python parent id": 1847, "Python id": 1848, "finished": true, "Ev Idx": 4748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901327.222, "dur": 0.566,
    "args": {
      "Python parent id": 12, "Python id": 1849, "finished": true, "Ev Idx": 4749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901327.418, "dur": 0.318,
    "args": {
      "Python parent id": 1849, "Python id": 1850, "finished": true, "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901328.357, "dur": 0.566,
    "args": {
      "Python parent id": 12, "Python id": 1851, "finished": true, "Ev Idx": 4751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901328.554, "dur": 0.316,
    "args": {
      "Python parent id": 1851, "Python id": 1852, "finished": true, "Ev Idx": 4752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901329.501, "dur": 0.569,
    "args": {
      "Python parent id": 12, "Python id": 1853, "finished": true, "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901329.718, "dur": 0.320,
    "args": {
      "Python parent id": 1853, "Python id": 1854, "finished": true, "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901330.638, "dur": 0.566,
    "args": {
      "Python parent id": 12, "Python id": 1855, "finished": true, "Ev Idx": 4755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901330.857, "dur": 0.317,
    "args": {
      "Python parent id": 1855, "Python id": 1856, "finished": true, "Ev Idx": 4756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901331.814, "dur": 0.530,
    "args": {
      "Python parent id": 12, "Python id": 1857, "finished": true, "Ev Idx": 4757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901332.006, "dur": 0.306,
    "args": {
      "Python parent id": 1857, "Python id": 1858, "finished": true, "Ev Idx": 4758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901332.918, "dur": 0.530,
    "args": {
      "Python parent id": 12, "Python id": 1859, "finished": true, "Ev Idx": 4759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901333.111, "dur": 0.303,
    "args": {
      "Python parent id": 1859, "Python id": 1860, "finished": true, "Ev Idx": 4760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901334.000, "dur": 0.596,
    "args": {
      "Python parent id": 12, "Python id": 1861, "finished": true, "Ev Idx": 4761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901334.217, "dur": 0.325,
    "args": {
      "Python parent id": 1861, "Python id": 1862, "finished": true, "Ev Idx": 4762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901335.156, "dur": 0.540,
    "args": {
      "Python parent id": 12, "Python id": 1863, "finished": true, "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901335.351, "dur": 0.311,
    "args": {
      "Python parent id": 1863, "Python id": 1864, "finished": true, "Ev Idx": 4764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901336.661, "dur": 0.551,
    "args": {
      "Python parent id": 12, "Python id": 1865, "finished": true, "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901336.858, "dur": 0.318,
    "args": {
      "Python parent id": 1865, "Python id": 1866, "finished": true, "Ev Idx": 4766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901337.764, "dur": 0.593,
    "args": {
      "Python parent id": 12, "Python id": 1867, "finished": true, "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901337.984, "dur": 0.321,
    "args": {
      "Python parent id": 1867, "Python id": 1868, "finished": true, "Ev Idx": 4768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901338.958, "dur": 2.144,
    "args": {
      "Python parent id": 12, "Python id": 1869, "finished": true, "Ev Idx": 4769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901339.178, "dur": 1.884,
    "args": {
      "Python parent id": 1869, "Python id": 1870, "finished": true, "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901341.705, "dur": 0.656,
    "args": {
      "Python parent id": 12, "Python id": 1871, "finished": true, "Ev Idx": 4771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901341.952, "dur": 0.375,
    "args": {
      "Python parent id": 1871, "Python id": 1872, "finished": true, "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901342.957, "dur": 0.599,
    "args": {
      "Python parent id": 12, "Python id": 1873, "finished": true, "Ev Idx": 4773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901343.180, "dur": 0.340,
    "args": {
      "Python parent id": 1873, "Python id": 1874, "finished": true, "Ev Idx": 4774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901344.138, "dur": 0.550,
    "args": {
      "Python parent id": 12, "Python id": 1875, "finished": true, "Ev Idx": 4775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901344.330, "dur": 0.325,
    "args": {
      "Python parent id": 1875, "Python id": 1876, "finished": true, "Ev Idx": 4776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(166): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901345.280, "dur": 0.601,
    "args": {
      "Python parent id": 12, "Python id": 1877, "finished": true, "Ev Idx": 4777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901345.507, "dur": 0.320,
    "args": {
      "Python parent id": 1877, "Python id": 1878, "finished": true, "Ev Idx": 4778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1030): synchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740901347.077, "dur": 36.766,
    "args": {
      "Python parent id": 12, "Python id": 1879, "finished": true, "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901347.503, "dur": 1.150,
    "args": {
      "Python parent id": 1879, "Python id": 1880, "finished": true, "Ev Idx": 4780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901347.839, "dur": 0.690,
    "args": {
      "Python parent id": 1880, "Python id": 1881, "finished": true, "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901348.238, "dur": 0.205,
    "args": {
      "Python parent id": 1881, "Python id": 1882, "finished": true, "Ev Idx": 4782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(490): __init__", "pid": 49625, "tid": 49625,
    "ts": 7253740901349.688, "dur": 24.417,
    "args": {
      "Python parent id": 1879, "Python id": 1883, "finished": true, "Ev Idx": 4783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901350.304, "dur": 23.371,
    "args": {
      "Python parent id": 1883, "Python id": 1884, "finished": true, "Ev Idx": 4784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901350.596, "dur": 0.303,
    "args": {
      "Python parent id": 1884, "Python id": 1885, "finished": true, "Ev Idx": 4785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901351.175, "dur": 0.065,
    "args": {
      "Python parent id": 1884, "Python id": 1886, "finished": true, "Ev Idx": 4786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901351.602, "dur": 0.058,
    "args": {
      "Python parent id": 1884, "Python id": 1887, "finished": true, "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901352.581, "dur": 0.093,
    "args": {
      "Python parent id": 1884, "Python id": 1888, "finished": true, "Ev Idx": 4788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901353.152, "dur": 0.159,
    "args": {
      "Python parent id": 1884, "Python id": 1889, "finished": true, "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901353.981, "dur": 19.557,
    "args": {
      "Python parent id": 1884, "Python id": 1890, "finished": true, "Ev Idx": 4790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901354.152, "dur": 0.055,
    "args": {
      "Python parent id": 1890, "Python id": 1891, "finished": true, "Ev Idx": 4791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901354.628, "dur": 0.054,
    "args": {
      "Python parent id": 1890, "Python id": 1892, "finished": true, "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901355.173, "dur": 0.130,
    "args": {
      "Python parent id": 1890, "Python id": 1893, "finished": true, "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901356.280, "dur": 0.474,
    "args": {
      "Python parent id": 1890, "Python id": 1894, "finished": true, "Ev Idx": 4794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901357.915, "dur": 15.424,
    "args": {
      "Python parent id": 1890, "Python id": 1895, "finished": true, "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740901358.447, "dur": 14.740,
    "args": {
      "Python parent id": 1895, "Python id": 1896, "finished": true, "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740901358.742, "dur": 9.727,
    "args": {
      "Python parent id": 1896, "Python id": 1897, "finished": true, "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740901359.192, "dur": 9.179,
    "args": {
      "Python parent id": 1897, "Python id": 1898, "finished": true, "Ev Idx": 4798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740901359.600, "dur": 0.795,
    "args": {
      "Python parent id": 1898, "Python id": 1899, "finished": true, "Ev Idx": 4799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740901359.999, "dur": 0.293,
    "args": {
      "Python parent id": 1899, "Python id": 1900, "finished": true, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740901360.949, "dur": 6.606,
    "args": {
      "Python parent id": 1898, "Python id": 1901, "finished": true, "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740901361.428, "dur": 5.927,
    "args": {
      "Python parent id": 1901, "Python id": 1902, "finished": true, "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740901361.915, "dur": 5.320,
    "args": {
      "Python parent id": 1902, "Python id": 1903, "finished": true, "Ev Idx": 4803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740901362.540, "dur": 3.731,
    "args": {
      "Python parent id": 1903, "Python id": 1904, "finished": true, "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740901363.017, "dur": 1.225,
    "args": {
      "Python parent id": 1904, "Python id": 1905, "finished": true, "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901363.215, "dur": 0.056,
    "args": {
      "Python parent id": 1905, "Python id": 1906, "finished": true, "Ev Idx": 4806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740901363.746, "dur": 0.373,
    "args": {
      "Python parent id": 1905, "Python id": 1907, "finished": true, "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740901367.974, "dur": 0.281,
    "args": {
      "Python parent id": 1898, "Python id": 1908, "finished": true, "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740901368.893, "dur": 0.288,
    "args": {
      "Python parent id": 1896, "Python id": 1909, "finished": true, "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740901370.087, "dur": 3.029,
    "args": {
      "Python parent id": 1896, "Python id": 1910, "finished": true, "Ev Idx": 4810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740901370.493, "dur": 2.560,
    "args": {
      "Python parent id": 1910, "Python id": 1911, "finished": true, "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901370.779, "dur": 0.611,
    "args": {
      "Python parent id": 1911, "Python id": 1912, "finished": true, "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901370.958, "dur": 0.333,
    "args": {
      "Python parent id": 1912, "Python id": 1913, "finished": true, "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901371.155, "dur": 0.063,
    "args": {
      "Python parent id": 1913, "Python id": 1914, "finished": true, "Ev Idx": 4814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740901372.012, "dur": 0.980,
    "args": {
      "Python parent id": 1911, "Python id": 1915, "finished": true, "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(494): __enter__", "pid": 49625, "tid": 49625,
    "ts": 7253740901374.918, "dur": 1.088,
    "args": {
      "Python parent id": 1879, "Python id": 1916, "finished": true, "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_exchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740901375.498, "dur": 0.364,
    "args": {
      "Python parent id": 1916, "Python id": 1917, "finished": true, "Ev Idx": 4817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_synchronize>", "pid": 49625, "tid": 49625,
    "ts": 7253740901376.567, "dur": 5.576,
    "args": {
      "Python parent id": 1879, "Python id": 1918, "finished": true, "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(497): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740901382.679, "dur": 0.832,
    "args": {
      "Python parent id": 1879, "Python id": 1919, "finished": true, "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_maybeExchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740901382.951, "dur": 0.365,
    "args": {
      "Python parent id": 1919, "Python id": 1920, "finished": true, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method empty of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740901389.602, "dur": 244.110,
    "args": {
      "Python parent id": 12, "Python id": 1921, "finished": true, "Ev Idx": 4821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740901639.450, "dur": 33.990,
    "args": {
      "Python parent id": 12, "Python id": 1922, "finished": true, "Ev Idx": 4822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740901674.656, "dur": 0.240,
    "args": {
      "Python parent id": 12, "Python id": 1923, "finished": true, "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901676.930, "dur": 0.675,
    "args": {
      "Python parent id": 12, "Python id": 1924, "finished": true, "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901678.867, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 1925, "finished": true, "Ev Idx": 4825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740901679.778, "dur": 26.987,
    "args": {
      "Python parent id": 12, "Python id": 1926, "finished": true, "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740901713.147, "dur": 55.637,
    "args": {
      "Python parent id": 12, "Python id": 1927, "finished": true, "Ev Idx": 4827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740901715.929, "dur": 40.302,
    "args": {
      "Python parent id": 1927, "Python id": 1928, "finished": true, "Ev Idx": 4828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901716.707, "dur": 1.216,
    "args": {
      "Python parent id": 1928, "Python id": 1929, "finished": true, "Ev Idx": 4829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901717.069, "dur": 0.673,
    "args": {
      "Python parent id": 1929, "Python id": 1930, "finished": true, "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901717.356, "dur": 0.210,
    "args": {
      "Python parent id": 1930, "Python id": 1931, "finished": true, "Ev Idx": 4831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901719.779, "dur": 19.210,
    "args": {
      "Python parent id": 1928, "Python id": 1932, "finished": true, "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901720.038, "dur": 0.116,
    "args": {
      "Python parent id": 1932, "Python id": 1933, "finished": true, "Ev Idx": 4833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901720.396, "dur": 0.046,
    "args": {
      "Python parent id": 1932, "Python id": 1934, "finished": true, "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901720.744, "dur": 0.048,
    "args": {
      "Python parent id": 1932, "Python id": 1935, "finished": true, "Ev Idx": 4835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901721.703, "dur": 0.049,
    "args": {
      "Python parent id": 1932, "Python id": 1936, "finished": true, "Ev Idx": 4836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901722.172, "dur": 0.145,
    "args": {
      "Python parent id": 1932, "Python id": 1937, "finished": true, "Ev Idx": 4837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901722.962, "dur": 15.869,
    "args": {
      "Python parent id": 1932, "Python id": 1938, "finished": true, "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901723.068, "dur": 0.105,
    "args": {
      "Python parent id": 1938, "Python id": 1939, "finished": true, "Ev Idx": 4839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901723.503, "dur": 0.059,
    "args": {
      "Python parent id": 1938, "Python id": 1940, "finished": true, "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901724.138, "dur": 0.077,
    "args": {
      "Python parent id": 1938, "Python id": 1941, "finished": true, "Ev Idx": 4841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901724.814, "dur": 0.031,
    "args": {
      "Python parent id": 1938, "Python id": 1942, "finished": true, "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901725.410, "dur": 13.238,
    "args": {
      "Python parent id": 1938, "Python id": 1943, "finished": true, "Ev Idx": 4843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740901725.972, "dur": 12.544,
    "args": {
      "Python parent id": 1943, "Python id": 1944, "finished": true, "Ev Idx": 4844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740901726.370, "dur": 8.971,
    "args": {
      "Python parent id": 1944, "Python id": 1945, "finished": true, "Ev Idx": 4845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740901726.829, "dur": 8.415,
    "args": {
      "Python parent id": 1945, "Python id": 1946, "finished": true, "Ev Idx": 4846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740901727.287, "dur": 0.579,
    "args": {
      "Python parent id": 1946, "Python id": 1947, "finished": true, "Ev Idx": 4847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740901727.572, "dur": 0.218,
    "args": {
      "Python parent id": 1947, "Python id": 1948, "finished": true, "Ev Idx": 4848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740901728.321, "dur": 6.212,
    "args": {
      "Python parent id": 1946, "Python id": 1949, "finished": true, "Ev Idx": 4849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740901728.925, "dur": 5.387,
    "args": {
      "Python parent id": 1949, "Python id": 1950, "finished": true, "Ev Idx": 4850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740901729.319, "dur": 4.906,
    "args": {
      "Python parent id": 1950, "Python id": 1951, "finished": true, "Ev Idx": 4851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740901729.754, "dur": 3.592,
    "args": {
      "Python parent id": 1951, "Python id": 1952, "finished": true, "Ev Idx": 4852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740901730.330, "dur": 1.262,
    "args": {
      "Python parent id": 1952, "Python id": 1953, "finished": true, "Ev Idx": 4853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901730.541, "dur": 0.049,
    "args": {
      "Python parent id": 1953, "Python id": 1954, "finished": true, "Ev Idx": 4854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740901731.081, "dur": 0.403,
    "args": {
      "Python parent id": 1953, "Python id": 1955, "finished": true, "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740901734.908, "dur": 0.267,
    "args": {
      "Python parent id": 1946, "Python id": 1956, "finished": true, "Ev Idx": 4856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740901735.720, "dur": 0.241,
    "args": {
      "Python parent id": 1944, "Python id": 1957, "finished": true, "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740901736.559, "dur": 1.899,
    "args": {
      "Python parent id": 1944, "Python id": 1958, "finished": true, "Ev Idx": 4858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740901736.855, "dur": 1.538,
    "args": {
      "Python parent id": 1958, "Python id": 1959, "finished": true, "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901737.098, "dur": 0.559,
    "args": {
      "Python parent id": 1959, "Python id": 1960, "finished": true, "Ev Idx": 4860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901737.329, "dur": 0.230,
    "args": {
      "Python parent id": 1960, "Python id": 1961, "finished": true, "Ev Idx": 4861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901737.436, "dur": 0.056,
    "args": {
      "Python parent id": 1961, "Python id": 1962, "finished": true, "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740901737.873, "dur": 0.460,
    "args": {
      "Python parent id": 1959, "Python id": 1963, "finished": true, "Ev Idx": 4863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740901740.286, "dur": 0.990,
    "args": {
      "Python parent id": 1928, "Python id": 1964, "finished": true, "Ev Idx": 4864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901751.336, "dur": 4.452,
    "args": {
      "Python parent id": 1928, "Python id": 1965, "finished": true, "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901753.421, "dur": 2.182,
    "args": {
      "Python parent id": 1965, "Python id": 1966, "finished": true, "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740901758.634, "dur": 9.842,
    "args": {
      "Python parent id": 1927, "Python id": 1967, "finished": true, "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740901774.539, "dur": 200.383,
    "args": {
      "Python parent id": 12, "Python id": 1968, "finished": true, "Ev Idx": 4868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740901775.573, "dur": 1.608,
    "args": {
      "Python parent id": 1968, "Python id": 1969, "finished": true, "Ev Idx": 4869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740901776.381, "dur": 0.698,
    "args": {
      "Python parent id": 1969, "Python id": 1970, "finished": true, "Ev Idx": 4870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740901777.943, "dur": 1.230,
    "args": {
      "Python parent id": 1968, "Python id": 1971, "finished": true, "Ev Idx": 4871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740901779.517, "dur": 0.279,
    "args": {
      "Python parent id": 1968, "Python id": 1972, "finished": true, "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740901780.605, "dur": 4.184,
    "args": {
      "Python parent id": 1968, "Python id": 1973, "finished": true, "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740901786.077, "dur": 188.645,
    "args": {
      "Python parent id": 1968, "Python id": 1974, "finished": true, "Ev Idx": 4874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740901789.343, "dur": 184.933,
    "args": {
      "Python parent id": 1974, "Python id": 1975, "finished": true, "Ev Idx": 4875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740901790.136, "dur": 184.055,
    "args": {
      "Python parent id": 1975, "Python id": 1976, "finished": true, "Ev Idx": 4876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740901807.709, "dur": 163.167,
    "args": {
      "Python parent id": 1976, "Python id": 1977, "finished": true, "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740901808.408, "dur": 162.304,
    "args": {
      "Python parent id": 1977, "Python id": 1978, "finished": true, "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740901810.629, "dur": 159.878,
    "args": {
      "Python parent id": 1978, "Python id": 1979, "finished": true, "Ev Idx": 4879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740901811.930, "dur": 0.561,
    "args": {
      "Python parent id": 1979, "Python id": 1980, "finished": true, "Ev Idx": 4880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740901812.266, "dur": 0.146,
    "args": {
      "Python parent id": 1980, "Python id": 1981, "finished": true, "Ev Idx": 4881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740901813.133, "dur": 0.401,
    "args": {
      "Python parent id": 1979, "Python id": 1982, "finished": true, "Ev Idx": 4882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740901813.384, "dur": 0.041,
    "args": {
      "Python parent id": 1982, "Python id": 1983, "finished": true, "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740901814.246, "dur": 156.009,
    "args": {
      "Python parent id": 1979, "Python id": 1984, "finished": true, "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740901815.980, "dur": 151.775,
    "args": {
      "Python parent id": 1984, "Python id": 1985, "finished": true, "Ev Idx": 4885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740901839.977, "dur": 49.490,
    "args": {
      "Python parent id": 1985, "Python id": 1986, "finished": true, "Ev Idx": 4886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740901840.585, "dur": 0.296,
    "args": {
      "Python parent id": 1986, "Python id": 1987, "finished": true, "Ev Idx": 4887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740901841.767, "dur": 47.501,
    "args": {
      "Python parent id": 1986, "Python id": 1988, "finished": true, "Ev Idx": 4888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740901842.315, "dur": 19.420,
    "args": {
      "Python parent id": 1988, "Python id": 1989, "finished": true, "Ev Idx": 4889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740901901.849, "dur": 12.714,
    "args": {
      "Python parent id": 1985, "Python id": 1990, "finished": true, "Ev Idx": 4890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740901930.163, "dur": 16.833,
    "args": {
      "Python parent id": 1985, "Python id": 1991, "finished": true, "Ev Idx": 4891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740901977.740, "dur": 27.707,
    "args": {
      "Python parent id": 12, "Python id": 1992, "finished": true, "Ev Idx": 4892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740901978.374, "dur": 21.936,
    "args": {
      "Python parent id": 1992, "Python id": 1993, "finished": true, "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901978.770, "dur": 0.848,
    "args": {
      "Python parent id": 1993, "Python id": 1994, "finished": true, "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901979.048, "dur": 0.458,
    "args": {
      "Python parent id": 1994, "Python id": 1995, "finished": true, "Ev Idx": 4895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901979.300, "dur": 0.111,
    "args": {
      "Python parent id": 1995, "Python id": 1996, "finished": true, "Ev Idx": 4896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901980.349, "dur": 15.037,
    "args": {
      "Python parent id": 1993, "Python id": 1997, "finished": true, "Ev Idx": 4897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901980.580, "dur": 0.138,
    "args": {
      "Python parent id": 1997, "Python id": 1998, "finished": true, "Ev Idx": 4898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901980.875, "dur": 0.077,
    "args": {
      "Python parent id": 1997, "Python id": 1999, "finished": true, "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901981.272, "dur": 0.061,
    "args": {
      "Python parent id": 1997, "Python id": 2000, "finished": true, "Ev Idx": 4900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901982.089, "dur": 0.052,
    "args": {
      "Python parent id": 1997, "Python id": 2001, "finished": true, "Ev Idx": 4901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901982.667, "dur": 0.088,
    "args": {
      "Python parent id": 1997, "Python id": 2002, "finished": true, "Ev Idx": 4902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901983.245, "dur": 12.070,
    "args": {
      "Python parent id": 1997, "Python id": 2003, "finished": true, "Ev Idx": 4903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901983.361, "dur": 0.063,
    "args": {
      "Python parent id": 2003, "Python id": 2004, "finished": true, "Ev Idx": 4904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901983.658, "dur": 0.053,
    "args": {
      "Python parent id": 2003, "Python id": 2005, "finished": true, "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901983.964, "dur": 0.082,
    "args": {
      "Python parent id": 2003, "Python id": 2006, "finished": true, "Ev Idx": 4906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740901984.571, "dur": 0.031,
    "args": {
      "Python parent id": 2003, "Python id": 2007, "finished": true, "Ev Idx": 4907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740901985.054, "dur": 10.105,
    "args": {
      "Python parent id": 2003, "Python id": 2008, "finished": true, "Ev Idx": 4908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740901985.471, "dur": 9.566,
    "args": {
      "Python parent id": 2008, "Python id": 2009, "finished": true, "Ev Idx": 4909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740901985.670, "dur": 6.531,
    "args": {
      "Python parent id": 2009, "Python id": 2010, "finished": true, "Ev Idx": 4910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740901985.981, "dur": 6.135,
    "args": {
      "Python parent id": 2010, "Python id": 2011, "finished": true, "Ev Idx": 4911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740901986.292, "dur": 0.488,
    "args": {
      "Python parent id": 2011, "Python id": 2012, "finished": true, "Ev Idx": 4912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740901986.532, "dur": 0.174,
    "args": {
      "Python parent id": 2012, "Python id": 2013, "finished": true, "Ev Idx": 4913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740901987.078, "dur": 4.268,
    "args": {
      "Python parent id": 2011, "Python id": 2014, "finished": true, "Ev Idx": 4914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740901987.470, "dur": 3.582,
    "args": {
      "Python parent id": 2014, "Python id": 2015, "finished": true, "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740901987.846, "dur": 3.077,
    "args": {
      "Python parent id": 2015, "Python id": 2016, "finished": true, "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740901988.129, "dur": 2.393,
    "args": {
      "Python parent id": 2016, "Python id": 2017, "finished": true, "Ev Idx": 4917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740901988.511, "dur": 0.972,
    "args": {
      "Python parent id": 2017, "Python id": 2018, "finished": true, "Ev Idx": 4918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740901988.695, "dur": 0.077,
    "args": {
      "Python parent id": 2018, "Python id": 2019, "finished": true, "Ev Idx": 4919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740901989.121, "dur": 0.254,
    "args": {
      "Python parent id": 2018, "Python id": 2020, "finished": true, "Ev Idx": 4920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740901991.730, "dur": 0.286,
    "args": {
      "Python parent id": 2011, "Python id": 2021, "finished": true, "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740901992.559, "dur": 0.177,
    "args": {
      "Python parent id": 2009, "Python id": 2022, "finished": true, "Ev Idx": 4922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740901993.316, "dur": 1.650,
    "args": {
      "Python parent id": 2009, "Python id": 2023, "finished": true, "Ev Idx": 4923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740901993.583, "dur": 1.293,
    "args": {
      "Python parent id": 2023, "Python id": 2024, "finished": true, "Ev Idx": 4924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740901993.839, "dur": 0.493,
    "args": {
      "Python parent id": 2024, "Python id": 2025, "finished": true, "Ev Idx": 4925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740901993.989, "dur": 0.257,
    "args": {
      "Python parent id": 2025, "Python id": 2026, "finished": true, "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740901994.122, "dur": 0.058,
    "args": {
      "Python parent id": 2026, "Python id": 2027, "finished": true, "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740901994.534, "dur": 0.271,
    "args": {
      "Python parent id": 2024, "Python id": 2028, "finished": true, "Ev Idx": 4928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740901995.711, "dur": 0.458,
    "args": {
      "Python parent id": 1993, "Python id": 2029, "finished": true, "Ev Idx": 4929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740901997.706, "dur": 2.309,
    "args": {
      "Python parent id": 1993, "Python id": 2030, "finished": true, "Ev Idx": 4930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740901998.913, "dur": 0.961,
    "args": {
      "Python parent id": 2030, "Python id": 2031, "finished": true, "Ev Idx": 4931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902001.027, "dur": 4.288,
    "args": {
      "Python parent id": 1992, "Python id": 2032, "finished": true, "Ev Idx": 4932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902006.556, "dur": 10.967,
    "args": {
      "Python parent id": 12, "Python id": 2033, "finished": true, "Ev Idx": 4933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902018.293, "dur": 0.178,
    "args": {
      "Python parent id": 12, "Python id": 2034, "finished": true, "Ev Idx": 4934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902019.436, "dur": 0.288,
    "args": {
      "Python parent id": 12, "Python id": 2035, "finished": true, "Ev Idx": 4935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902020.494, "dur": 0.078,
    "args": {
      "Python parent id": 12, "Python id": 2036, "finished": true, "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902021.176, "dur": 16.123,
    "args": {
      "Python parent id": 12, "Python id": 2037, "finished": true, "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902039.108, "dur": 27.933,
    "args": {
      "Python parent id": 12, "Python id": 2038, "finished": true, "Ev Idx": 4938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902039.557, "dur": 23.243,
    "args": {
      "Python parent id": 2038, "Python id": 2039, "finished": true, "Ev Idx": 4939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902039.779, "dur": 0.615,
    "args": {
      "Python parent id": 2039, "Python id": 2040, "finished": true, "Ev Idx": 4940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902039.975, "dur": 0.338,
    "args": {
      "Python parent id": 2040, "Python id": 2041, "finished": true, "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902040.185, "dur": 0.081,
    "args": {
      "Python parent id": 2041, "Python id": 2042, "finished": true, "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902040.869, "dur": 19.102,
    "args": {
      "Python parent id": 2039, "Python id": 2043, "finished": true, "Ev Idx": 4943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902040.968, "dur": 0.059,
    "args": {
      "Python parent id": 2043, "Python id": 2044, "finished": true, "Ev Idx": 4944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902041.233, "dur": 0.055,
    "args": {
      "Python parent id": 2043, "Python id": 2045, "finished": true, "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902041.558, "dur": 0.057,
    "args": {
      "Python parent id": 2043, "Python id": 2046, "finished": true, "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902042.127, "dur": 0.046,
    "args": {
      "Python parent id": 2043, "Python id": 2047, "finished": true, "Ev Idx": 4947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902042.597, "dur": 0.078,
    "args": {
      "Python parent id": 2043, "Python id": 2048, "finished": true, "Ev Idx": 4948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902043.122, "dur": 16.768,
    "args": {
      "Python parent id": 2043, "Python id": 2049, "finished": true, "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902043.211, "dur": 0.072,
    "args": {
      "Python parent id": 2049, "Python id": 2050, "finished": true, "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902043.501, "dur": 0.062,
    "args": {
      "Python parent id": 2049, "Python id": 2051, "finished": true, "Ev Idx": 4951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902043.806, "dur": 0.071,
    "args": {
      "Python parent id": 2049, "Python id": 2052, "finished": true, "Ev Idx": 4952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902044.355, "dur": 0.060,
    "args": {
      "Python parent id": 2049, "Python id": 2053, "finished": true, "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902044.803, "dur": 14.960,
    "args": {
      "Python parent id": 2049, "Python id": 2054, "finished": true, "Ev Idx": 4954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902045.107, "dur": 14.498,
    "args": {
      "Python parent id": 2054, "Python id": 2055, "finished": true, "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902045.260, "dur": 12.005,
    "args": {
      "Python parent id": 2055, "Python id": 2056, "finished": true, "Ev Idx": 4956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902045.520, "dur": 11.621,
    "args": {
      "Python parent id": 2056, "Python id": 2057, "finished": true, "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902045.664, "dur": 0.326,
    "args": {
      "Python parent id": 2057, "Python id": 2058, "finished": true, "Ev Idx": 4958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902045.815, "dur": 0.076,
    "args": {
      "Python parent id": 2058, "Python id": 2059, "finished": true, "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902046.245, "dur": 10.404,
    "args": {
      "Python parent id": 2057, "Python id": 2060, "finished": true, "Ev Idx": 4960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902046.531, "dur": 9.934,
    "args": {
      "Python parent id": 2060, "Python id": 2061, "finished": true, "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902046.716, "dur": 9.660,
    "args": {
      "Python parent id": 2061, "Python id": 2062, "finished": true, "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902046.899, "dur": 9.205,
    "args": {
      "Python parent id": 2062, "Python id": 2063, "finished": true, "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902047.178, "dur": 8.283,
    "args": {
      "Python parent id": 2063, "Python id": 2064, "finished": true, "Ev Idx": 4964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902047.292, "dur": 0.042,
    "args": {
      "Python parent id": 2064, "Python id": 2065, "finished": true, "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902047.558, "dur": 7.813,
    "args": {
      "Python parent id": 2064, "Python id": 2066, "finished": true, "Ev Idx": 4966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902056.977, "dur": 0.083,
    "args": {
      "Python parent id": 2057, "Python id": 2067, "finished": true, "Ev Idx": 4967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902057.462, "dur": 0.141,
    "args": {
      "Python parent id": 2055, "Python id": 2068, "finished": true, "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902058.017, "dur": 1.541,
    "args": {
      "Python parent id": 2055, "Python id": 2069, "finished": true, "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902058.235, "dur": 1.257,
    "args": {
      "Python parent id": 2069, "Python id": 2070, "finished": true, "Ev Idx": 4970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902058.396, "dur": 0.461,
    "args": {
      "Python parent id": 2070, "Python id": 2071, "finished": true, "Ev Idx": 4971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902058.568, "dur": 0.225,
    "args": {
      "Python parent id": 2071, "Python id": 2072, "finished": true, "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902058.682, "dur": 0.080,
    "args": {
      "Python parent id": 2072, "Python id": 2073, "finished": true, "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902059.142, "dur": 0.304,
    "args": {
      "Python parent id": 2070, "Python id": 2074, "finished": true, "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902060.181, "dur": 0.171,
    "args": {
      "Python parent id": 2039, "Python id": 2075, "finished": true, "Ev Idx": 4975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902061.267, "dur": 1.292,
    "args": {
      "Python parent id": 2039, "Python id": 2076, "finished": true, "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902061.906, "dur": 0.583,
    "args": {
      "Python parent id": 2076, "Python id": 2077, "finished": true, "Ev Idx": 4977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902063.269, "dur": 3.692,
    "args": {
      "Python parent id": 2038, "Python id": 2078, "finished": true, "Ev Idx": 4978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902068.287, "dur": 140.082,
    "args": {
      "Python parent id": 12, "Python id": 2079, "finished": true, "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902068.771, "dur": 13.946,
    "args": {
      "Python parent id": 2079, "Python id": 2080, "finished": true, "Ev Idx": 4980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902082.306, "dur": 0.323,
    "args": {
      "Python parent id": 2080, "Python id": 2081, "finished": true, "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902083.325, "dur": 0.654,
    "args": {
      "Python parent id": 2079, "Python id": 2082, "finished": true, "Ev Idx": 4982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902084.265, "dur": 0.280,
    "args": {
      "Python parent id": 2079, "Python id": 2083, "finished": true, "Ev Idx": 4983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902084.932, "dur": 4.090,
    "args": {
      "Python parent id": 2079, "Python id": 2084, "finished": true, "Ev Idx": 4984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902089.839, "dur": 118.363,
    "args": {
      "Python parent id": 2079, "Python id": 2085, "finished": true, "Ev Idx": 4985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902091.276, "dur": 116.571,
    "args": {
      "Python parent id": 2085, "Python id": 2086, "finished": true, "Ev Idx": 4986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902091.756, "dur": 116.011,
    "args": {
      "Python parent id": 2086, "Python id": 2087, "finished": true, "Ev Idx": 4987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902101.426, "dur": 104.025,
    "args": {
      "Python parent id": 2087, "Python id": 2088, "finished": true, "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740902101.836, "dur": 103.469,
    "args": {
      "Python parent id": 2088, "Python id": 2089, "finished": true, "Ev Idx": 4989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740902102.745, "dur": 102.374,
    "args": {
      "Python parent id": 2089, "Python id": 2090, "finished": true, "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902103.451, "dur": 0.490,
    "args": {
      "Python parent id": 2090, "Python id": 2091, "finished": true, "Ev Idx": 4991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902103.780, "dur": 0.105,
    "args": {
      "Python parent id": 2091, "Python id": 2092, "finished": true, "Ev Idx": 4992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902104.313, "dur": 0.342,
    "args": {
      "Python parent id": 2090, "Python id": 2093, "finished": true, "Ev Idx": 4993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902104.559, "dur": 0.040,
    "args": {
      "Python parent id": 2093, "Python id": 2094, "finished": true, "Ev Idx": 4994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902105.113, "dur": 99.809,
    "args": {
      "Python parent id": 2090, "Python id": 2095, "finished": true, "Ev Idx": 4995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902106.136, "dur": 97.093,
    "args": {
      "Python parent id": 2095, "Python id": 2096, "finished": true, "Ev Idx": 4996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740902117.581, "dur": 28.916,
    "args": {
      "Python parent id": 2096, "Python id": 2097, "finished": true, "Ev Idx": 4997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740902117.759, "dur": 0.213,
    "args": {
      "Python parent id": 2097, "Python id": 2098, "finished": true, "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740902118.468, "dur": 27.823,
    "args": {
      "Python parent id": 2097, "Python id": 2099, "finished": true, "Ev Idx": 4999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902118.653, "dur": 12.896,
    "args": {
      "Python parent id": 2099, "Python id": 2100, "finished": true, "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902156.432, "dur": 8.548,
    "args": {
      "Python parent id": 2096, "Python id": 2101, "finished": true, "Ev Idx": 5001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902175.305, "dur": 10.125,
    "args": {
      "Python parent id": 2096, "Python id": 2102, "finished": true, "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902210.512, "dur": 22.101,
    "args": {
      "Python parent id": 12, "Python id": 2103, "finished": true, "Ev Idx": 5003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902211.074, "dur": 16.867,
    "args": {
      "Python parent id": 2103, "Python id": 2104, "finished": true, "Ev Idx": 5004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902211.271, "dur": 0.670,
    "args": {
      "Python parent id": 2104, "Python id": 2105, "finished": true, "Ev Idx": 5005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902211.479, "dur": 0.358,
    "args": {
      "Python parent id": 2105, "Python id": 2106, "finished": true, "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902211.641, "dur": 0.140,
    "args": {
      "Python parent id": 2106, "Python id": 2107, "finished": true, "Ev Idx": 5007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902212.434, "dur": 11.941,
    "args": {
      "Python parent id": 2104, "Python id": 2108, "finished": true, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902212.627, "dur": 0.083,
    "args": {
      "Python parent id": 2108, "Python id": 2109, "finished": true, "Ev Idx": 5009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902212.858, "dur": 0.059,
    "args": {
      "Python parent id": 2108, "Python id": 2110, "finished": true, "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902213.165, "dur": 0.054,
    "args": {
      "Python parent id": 2108, "Python id": 2111, "finished": true, "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902213.858, "dur": 0.050,
    "args": {
      "Python parent id": 2108, "Python id": 2112, "finished": true, "Ev Idx": 5012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902214.185, "dur": 0.092,
    "args": {
      "Python parent id": 2108, "Python id": 2113, "finished": true, "Ev Idx": 5013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902214.678, "dur": 9.643,
    "args": {
      "Python parent id": 2108, "Python id": 2114, "finished": true, "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902214.777, "dur": 0.066,
    "args": {
      "Python parent id": 2114, "Python id": 2115, "finished": true, "Ev Idx": 5015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902215.080, "dur": 0.058,
    "args": {
      "Python parent id": 2114, "Python id": 2116, "finished": true, "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902215.369, "dur": 0.077,
    "args": {
      "Python parent id": 2114, "Python id": 2117, "finished": true, "Ev Idx": 5017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902215.793, "dur": 0.062,
    "args": {
      "Python parent id": 2114, "Python id": 2118, "finished": true, "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902216.313, "dur": 7.888,
    "args": {
      "Python parent id": 2114, "Python id": 2119, "finished": true, "Ev Idx": 5019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902216.591, "dur": 7.518,
    "args": {
      "Python parent id": 2119, "Python id": 2120, "finished": true, "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902216.719, "dur": 4.916,
    "args": {
      "Python parent id": 2120, "Python id": 2121, "finished": true, "Ev Idx": 5021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902216.932, "dur": 4.609,
    "args": {
      "Python parent id": 2121, "Python id": 2122, "finished": true, "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902217.101, "dur": 0.399,
    "args": {
      "Python parent id": 2122, "Python id": 2123, "finished": true, "Ev Idx": 5023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902217.321, "dur": 0.121,
    "args": {
      "Python parent id": 2123, "Python id": 2124, "finished": true, "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902217.725, "dur": 3.270,
    "args": {
      "Python parent id": 2122, "Python id": 2125, "finished": true, "Ev Idx": 5025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902218.075, "dur": 2.732,
    "args": {
      "Python parent id": 2125, "Python id": 2126, "finished": true, "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902218.358, "dur": 2.359,
    "args": {
      "Python parent id": 2126, "Python id": 2127, "finished": true, "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902218.562, "dur": 1.789,
    "args": {
      "Python parent id": 2127, "Python id": 2128, "finished": true, "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902218.834, "dur": 0.716,
    "args": {
      "Python parent id": 2128, "Python id": 2129, "finished": true, "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902218.970, "dur": 0.044,
    "args": {
      "Python parent id": 2129, "Python id": 2130, "finished": true, "Ev Idx": 5030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902219.243, "dur": 0.239,
    "args": {
      "Python parent id": 2129, "Python id": 2131, "finished": true, "Ev Idx": 5031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902221.255, "dur": 0.196,
    "args": {
      "Python parent id": 2122, "Python id": 2132, "finished": true, "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902221.856, "dur": 0.171,
    "args": {
      "Python parent id": 2120, "Python id": 2133, "finished": true, "Ev Idx": 5033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902222.502, "dur": 1.563,
    "args": {
      "Python parent id": 2120, "Python id": 2134, "finished": true, "Ev Idx": 5034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902222.773, "dur": 1.217,
    "args": {
      "Python parent id": 2134, "Python id": 2135, "finished": true, "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902222.925, "dur": 0.500,
    "args": {
      "Python parent id": 2135, "Python id": 2136, "finished": true, "Ev Idx": 5036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902223.119, "dur": 0.206,
    "args": {
      "Python parent id": 2136, "Python id": 2137, "finished": true, "Ev Idx": 5037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902223.217, "dur": 0.061,
    "args": {
      "Python parent id": 2137, "Python id": 2138, "finished": true, "Ev Idx": 5038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902223.577, "dur": 0.368,
    "args": {
      "Python parent id": 2135, "Python id": 2139, "finished": true, "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902224.584, "dur": 0.350,
    "args": {
      "Python parent id": 2104, "Python id": 2140, "finished": true, "Ev Idx": 5040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902225.898, "dur": 1.761,
    "args": {
      "Python parent id": 2104, "Python id": 2141, "finished": true, "Ev Idx": 5041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902226.741, "dur": 0.808,
    "args": {
      "Python parent id": 2141, "Python id": 2142, "finished": true, "Ev Idx": 5042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902228.446, "dur": 4.045,
    "args": {
      "Python parent id": 2103, "Python id": 2143, "finished": true, "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902233.331, "dur": 9.194,
    "args": {
      "Python parent id": 12, "Python id": 2144, "finished": true, "Ev Idx": 5044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902243.074, "dur": 0.171,
    "args": {
      "Python parent id": 12, "Python id": 2145, "finished": true, "Ev Idx": 5045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902244.089, "dur": 0.279,
    "args": {
      "Python parent id": 12, "Python id": 2146, "finished": true, "Ev Idx": 5046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902244.977, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 2147, "finished": true, "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902245.570, "dur": 12.107,
    "args": {
      "Python parent id": 12, "Python id": 2148, "finished": true, "Ev Idx": 5048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902259.383, "dur": 18.187,
    "args": {
      "Python parent id": 12, "Python id": 2149, "finished": true, "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902259.826, "dur": 13.391,
    "args": {
      "Python parent id": 2149, "Python id": 2150, "finished": true, "Ev Idx": 5050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902259.987, "dur": 0.520,
    "args": {
      "Python parent id": 2150, "Python id": 2151, "finished": true, "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902260.152, "dur": 0.287,
    "args": {
      "Python parent id": 2151, "Python id": 2152, "finished": true, "Ev Idx": 5052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902260.308, "dur": 0.090,
    "args": {
      "Python parent id": 2152, "Python id": 2153, "finished": true, "Ev Idx": 5053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902260.965, "dur": 10.080,
    "args": {
      "Python parent id": 2150, "Python id": 2154, "finished": true, "Ev Idx": 5054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902261.060, "dur": 0.072,
    "args": {
      "Python parent id": 2154, "Python id": 2155, "finished": true, "Ev Idx": 5055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902261.319, "dur": 0.063,
    "args": {
      "Python parent id": 2154, "Python id": 2156, "finished": true, "Ev Idx": 5056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902261.554, "dur": 0.053,
    "args": {
      "Python parent id": 2154, "Python id": 2157, "finished": true, "Ev Idx": 5057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902262.033, "dur": 0.059,
    "args": {
      "Python parent id": 2154, "Python id": 2158, "finished": true, "Ev Idx": 5058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902262.403, "dur": 0.071,
    "args": {
      "Python parent id": 2154, "Python id": 2159, "finished": true, "Ev Idx": 5059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902262.934, "dur": 8.050,
    "args": {
      "Python parent id": 2154, "Python id": 2160, "finished": true, "Ev Idx": 5060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902263.026, "dur": 0.061,
    "args": {
      "Python parent id": 2160, "Python id": 2161, "finished": true, "Ev Idx": 5061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902263.243, "dur": 0.060,
    "args": {
      "Python parent id": 2160, "Python id": 2162, "finished": true, "Ev Idx": 5062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902263.488, "dur": 0.084,
    "args": {
      "Python parent id": 2160, "Python id": 2163, "finished": true, "Ev Idx": 5063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902263.911, "dur": 0.054,
    "args": {
      "Python parent id": 2160, "Python id": 2164, "finished": true, "Ev Idx": 5064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902264.322, "dur": 6.532,
    "args": {
      "Python parent id": 2160, "Python id": 2165, "finished": true, "Ev Idx": 5065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902264.605, "dur": 6.170,
    "args": {
      "Python parent id": 2165, "Python id": 2166, "finished": true, "Ev Idx": 5066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902264.757, "dur": 4.026,
    "args": {
      "Python parent id": 2166, "Python id": 2167, "finished": true, "Ev Idx": 5067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902265.020, "dur": 3.690,
    "args": {
      "Python parent id": 2167, "Python id": 2168, "finished": true, "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902265.236, "dur": 0.406,
    "args": {
      "Python parent id": 2168, "Python id": 2169, "finished": true, "Ev Idx": 5069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902265.490, "dur": 0.113,
    "args": {
      "Python parent id": 2169, "Python id": 2170, "finished": true, "Ev Idx": 5070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902265.855, "dur": 2.503,
    "args": {
      "Python parent id": 2168, "Python id": 2171, "finished": true, "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902266.117, "dur": 2.164,
    "args": {
      "Python parent id": 2171, "Python id": 2172, "finished": true, "Ev Idx": 5072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902266.356, "dur": 1.845,
    "args": {
      "Python parent id": 2172, "Python id": 2173, "finished": true, "Ev Idx": 5073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902266.551, "dur": 1.391,
    "args": {
      "Python parent id": 2173, "Python id": 2174, "finished": true, "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902266.787, "dur": 0.601,
    "args": {
      "Python parent id": 2174, "Python id": 2175, "finished": true, "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902266.896, "dur": 0.078,
    "args": {
      "Python parent id": 2175, "Python id": 2176, "finished": true, "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902267.122, "dur": 0.200,
    "args": {
      "Python parent id": 2175, "Python id": 2177, "finished": true, "Ev Idx": 5077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902268.571, "dur": 0.085,
    "args": {
      "Python parent id": 2168, "Python id": 2178, "finished": true, "Ev Idx": 5078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902268.960, "dur": 0.181,
    "args": {
      "Python parent id": 2166, "Python id": 2179, "finished": true, "Ev Idx": 5079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902269.477, "dur": 1.232,
    "args": {
      "Python parent id": 2166, "Python id": 2180, "finished": true, "Ev Idx": 5080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902269.649, "dur": 1.004,
    "args": {
      "Python parent id": 2180, "Python id": 2181, "finished": true, "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902269.801, "dur": 0.407,
    "args": {
      "Python parent id": 2181, "Python id": 2182, "finished": true, "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902269.938, "dur": 0.207,
    "args": {
      "Python parent id": 2182, "Python id": 2183, "finished": true, "Ev Idx": 5083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902270.042, "dur": 0.073,
    "args": {
      "Python parent id": 2183, "Python id": 2184, "finished": true, "Ev Idx": 5084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902270.361, "dur": 0.250,
    "args": {
      "Python parent id": 2181, "Python id": 2185, "finished": true, "Ev Idx": 5085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902271.177, "dur": 0.213,
    "args": {
      "Python parent id": 2150, "Python id": 2186, "finished": true, "Ev Idx": 5086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902272.058, "dur": 1.021,
    "args": {
      "Python parent id": 2150, "Python id": 2187, "finished": true, "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902272.500, "dur": 0.529,
    "args": {
      "Python parent id": 2187, "Python id": 2188, "finished": true, "Ev Idx": 5088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902273.545, "dur": 3.981,
    "args": {
      "Python parent id": 2149, "Python id": 2189, "finished": true, "Ev Idx": 5089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902278.622, "dur": 115.780,
    "args": {
      "Python parent id": 12, "Python id": 2190, "finished": true, "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902279.032, "dur": 0.548,
    "args": {
      "Python parent id": 2190, "Python id": 2191, "finished": true, "Ev Idx": 5091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902279.427, "dur": 0.107,
    "args": {
      "Python parent id": 2191, "Python id": 2192, "finished": true, "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902280.047, "dur": 0.375,
    "args": {
      "Python parent id": 2190, "Python id": 2193, "finished": true, "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902280.680, "dur": 0.238,
    "args": {
      "Python parent id": 2190, "Python id": 2194, "finished": true, "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902281.187, "dur": 1.939,
    "args": {
      "Python parent id": 2190, "Python id": 2195, "finished": true, "Ev Idx": 5095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902283.785, "dur": 110.503,
    "args": {
      "Python parent id": 2190, "Python id": 2196, "finished": true, "Ev Idx": 5096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902284.837, "dur": 109.224,
    "args": {
      "Python parent id": 2196, "Python id": 2197, "finished": true, "Ev Idx": 5097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902285.208, "dur": 108.789,
    "args": {
      "Python parent id": 2197, "Python id": 2198, "finished": true, "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902295.336, "dur": 96.929,
    "args": {
      "Python parent id": 2198, "Python id": 2199, "finished": true, "Ev Idx": 5099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740902295.702, "dur": 96.510,
    "args": {
      "Python parent id": 2199, "Python id": 2200, "finished": true, "Ev Idx": 5100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740902296.439, "dur": 95.647,
    "args": {
      "Python parent id": 2200, "Python id": 2201, "finished": true, "Ev Idx": 5101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902297.090, "dur": 0.528,
    "args": {
      "Python parent id": 2201, "Python id": 2202, "finished": true, "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902297.455, "dur": 0.095,
    "args": {
      "Python parent id": 2202, "Python id": 2203, "finished": true, "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902298.037, "dur": 0.390,
    "args": {
      "Python parent id": 2201, "Python id": 2204, "finished": true, "Ev Idx": 5104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902298.327, "dur": 0.044,
    "args": {
      "Python parent id": 2204, "Python id": 2205, "finished": true, "Ev Idx": 5105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902298.797, "dur": 93.163,
    "args": {
      "Python parent id": 2201, "Python id": 2206, "finished": true, "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902299.726, "dur": 90.732,
    "args": {
      "Python parent id": 2206, "Python id": 2207, "finished": true, "Ev Idx": 5107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740902309.247, "dur": 26.069,
    "args": {
      "Python parent id": 2207, "Python id": 2208, "finished": true, "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740902309.479, "dur": 0.137,
    "args": {
      "Python parent id": 2208, "Python id": 2209, "finished": true, "Ev Idx": 5109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740902310.066, "dur": 25.080,
    "args": {
      "Python parent id": 2208, "Python id": 2210, "finished": true, "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902310.228, "dur": 9.643,
    "args": {
      "Python parent id": 2210, "Python id": 2211, "finished": true, "Ev Idx": 5111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902344.647, "dur": 7.782,
    "args": {
      "Python parent id": 2207, "Python id": 2212, "finished": true, "Ev Idx": 5112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902362.129, "dur": 10.506,
    "args": {
      "Python parent id": 2207, "Python id": 2213, "finished": true, "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902396.867, "dur": 18.048,
    "args": {
      "Python parent id": 12, "Python id": 2214, "finished": true, "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902397.393, "dur": 14.204,
    "args": {
      "Python parent id": 2214, "Python id": 2215, "finished": true, "Ev Idx": 5115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902397.604, "dur": 0.574,
    "args": {
      "Python parent id": 2215, "Python id": 2216, "finished": true, "Ev Idx": 5116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902397.828, "dur": 0.283,
    "args": {
      "Python parent id": 2216, "Python id": 2217, "finished": true, "Ev Idx": 5117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902397.951, "dur": 0.092,
    "args": {
      "Python parent id": 2217, "Python id": 2218, "finished": true, "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902398.592, "dur": 10.348,
    "args": {
      "Python parent id": 2215, "Python id": 2219, "finished": true, "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902398.716, "dur": 0.106,
    "args": {
      "Python parent id": 2219, "Python id": 2220, "finished": true, "Ev Idx": 5120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902398.957, "dur": 0.060,
    "args": {
      "Python parent id": 2219, "Python id": 2221, "finished": true, "Ev Idx": 5121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902399.221, "dur": 0.051,
    "args": {
      "Python parent id": 2219, "Python id": 2222, "finished": true, "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902399.808, "dur": 0.040,
    "args": {
      "Python parent id": 2219, "Python id": 2223, "finished": true, "Ev Idx": 5123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902400.170, "dur": 0.092,
    "args": {
      "Python parent id": 2219, "Python id": 2224, "finished": true, "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902400.629, "dur": 8.254,
    "args": {
      "Python parent id": 2219, "Python id": 2225, "finished": true, "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902400.735, "dur": 0.062,
    "args": {
      "Python parent id": 2225, "Python id": 2226, "finished": true, "Ev Idx": 5126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902400.977, "dur": 0.063,
    "args": {
      "Python parent id": 2225, "Python id": 2227, "finished": true, "Ev Idx": 5127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902401.257, "dur": 0.076,
    "args": {
      "Python parent id": 2225, "Python id": 2228, "finished": true, "Ev Idx": 5128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902401.744, "dur": 0.052,
    "args": {
      "Python parent id": 2225, "Python id": 2229, "finished": true, "Ev Idx": 5129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902402.113, "dur": 6.663,
    "args": {
      "Python parent id": 2225, "Python id": 2230, "finished": true, "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902402.399, "dur": 6.320,
    "args": {
      "Python parent id": 2230, "Python id": 2231, "finished": true, "Ev Idx": 5131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902402.564, "dur": 4.058,
    "args": {
      "Python parent id": 2231, "Python id": 2232, "finished": true, "Ev Idx": 5132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902402.776, "dur": 3.751,
    "args": {
      "Python parent id": 2232, "Python id": 2233, "finished": true, "Ev Idx": 5133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902402.975, "dur": 0.306,
    "args": {
      "Python parent id": 2233, "Python id": 2234, "finished": true, "Ev Idx": 5134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902403.150, "dur": 0.068,
    "args": {
      "Python parent id": 2234, "Python id": 2235, "finished": true, "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902403.500, "dur": 2.616,
    "args": {
      "Python parent id": 2233, "Python id": 2236, "finished": true, "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902403.785, "dur": 2.194,
    "args": {
      "Python parent id": 2236, "Python id": 2237, "finished": true, "Ev Idx": 5137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902404.028, "dur": 1.871,
    "args": {
      "Python parent id": 2237, "Python id": 2238, "finished": true, "Ev Idx": 5138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902404.249, "dur": 1.432,
    "args": {
      "Python parent id": 2238, "Python id": 2239, "finished": true, "Ev Idx": 5139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902404.539, "dur": 0.583,
    "args": {
      "Python parent id": 2239, "Python id": 2240, "finished": true, "Ev Idx": 5140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902404.634, "dur": 0.047,
    "args": {
      "Python parent id": 2240, "Python id": 2241, "finished": true, "Ev Idx": 5141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902404.866, "dur": 0.174,
    "args": {
      "Python parent id": 2240, "Python id": 2242, "finished": true, "Ev Idx": 5142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902406.321, "dur": 0.163,
    "args": {
      "Python parent id": 2233, "Python id": 2243, "finished": true, "Ev Idx": 5143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902406.783, "dur": 0.145,
    "args": {
      "Python parent id": 2231, "Python id": 2244, "finished": true, "Ev Idx": 5144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902407.255, "dur": 1.402,
    "args": {
      "Python parent id": 2231, "Python id": 2245, "finished": true, "Ev Idx": 5145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902407.476, "dur": 1.120,
    "args": {
      "Python parent id": 2245, "Python id": 2246, "finished": true, "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902407.625, "dur": 0.454,
    "args": {
      "Python parent id": 2246, "Python id": 2247, "finished": true, "Ev Idx": 5147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902407.803, "dur": 0.208,
    "args": {
      "Python parent id": 2247, "Python id": 2248, "finished": true, "Ev Idx": 5148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902407.902, "dur": 0.076,
    "args": {
      "Python parent id": 2248, "Python id": 2249, "finished": true, "Ev Idx": 5149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902408.228, "dur": 0.324,
    "args": {
      "Python parent id": 2246, "Python id": 2250, "finished": true, "Ev Idx": 5150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902409.080, "dur": 0.277,
    "args": {
      "Python parent id": 2215, "Python id": 2251, "finished": true, "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902410.179, "dur": 1.257,
    "args": {
      "Python parent id": 2215, "Python id": 2252, "finished": true, "Ev Idx": 5152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902410.714, "dur": 0.670,
    "args": {
      "Python parent id": 2252, "Python id": 2253, "finished": true, "Ev Idx": 5153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902411.985, "dur": 2.850,
    "args": {
      "Python parent id": 2214, "Python id": 2254, "finished": true, "Ev Idx": 5154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902415.539, "dur": 7.656,
    "args": {
      "Python parent id": 12, "Python id": 2255, "finished": true, "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902423.668, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 2256, "finished": true, "Ev Idx": 5156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902424.547, "dur": 0.218,
    "args": {
      "Python parent id": 12, "Python id": 2257, "finished": true, "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902425.288, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 2258, "finished": true, "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902425.795, "dur": 10.694,
    "args": {
      "Python parent id": 12, "Python id": 2259, "finished": true, "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902437.990, "dur": 14.858,
    "args": {
      "Python parent id": 12, "Python id": 2260, "finished": true, "Ev Idx": 5160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902438.345, "dur": 12.060,
    "args": {
      "Python parent id": 2260, "Python id": 2261, "finished": true, "Ev Idx": 5161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902438.552, "dur": 0.545,
    "args": {
      "Python parent id": 2261, "Python id": 2262, "finished": true, "Ev Idx": 5162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902438.756, "dur": 0.272,
    "args": {
      "Python parent id": 2262, "Python id": 2263, "finished": true, "Ev Idx": 5163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902438.876, "dur": 0.116,
    "args": {
      "Python parent id": 2263, "Python id": 2264, "finished": true, "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902439.371, "dur": 8.943,
    "args": {
      "Python parent id": 2261, "Python id": 2265, "finished": true, "Ev Idx": 5165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902439.503, "dur": 0.070,
    "args": {
      "Python parent id": 2265, "Python id": 2266, "finished": true, "Ev Idx": 5166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902439.727, "dur": 0.060,
    "args": {
      "Python parent id": 2265, "Python id": 2267, "finished": true, "Ev Idx": 5167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902439.941, "dur": 0.056,
    "args": {
      "Python parent id": 2265, "Python id": 2268, "finished": true, "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902440.402, "dur": 0.041,
    "args": {
      "Python parent id": 2265, "Python id": 2269, "finished": true, "Ev Idx": 5169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902440.685, "dur": 0.081,
    "args": {
      "Python parent id": 2265, "Python id": 2270, "finished": true, "Ev Idx": 5170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902441.166, "dur": 7.076,
    "args": {
      "Python parent id": 2265, "Python id": 2271, "finished": true, "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902441.272, "dur": 0.069,
    "args": {
      "Python parent id": 2271, "Python id": 2272, "finished": true, "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902441.505, "dur": 0.053,
    "args": {
      "Python parent id": 2271, "Python id": 2273, "finished": true, "Ev Idx": 5173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902441.760, "dur": 0.072,
    "args": {
      "Python parent id": 2271, "Python id": 2274, "finished": true, "Ev Idx": 5174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902442.204, "dur": 0.029,
    "args": {
      "Python parent id": 2271, "Python id": 2275, "finished": true, "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902442.566, "dur": 5.597,
    "args": {
      "Python parent id": 2271, "Python id": 2276, "finished": true, "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902442.898, "dur": 5.175,
    "args": {
      "Python parent id": 2276, "Python id": 2277, "finished": true, "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902443.023, "dur": 3.342,
    "args": {
      "Python parent id": 2277, "Python id": 2278, "finished": true, "Ev Idx": 5178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902443.202, "dur": 3.090,
    "args": {
      "Python parent id": 2278, "Python id": 2279, "finished": true, "Ev Idx": 5179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902443.368, "dur": 0.315,
    "args": {
      "Python parent id": 2279, "Python id": 2280, "finished": true, "Ev Idx": 5180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902443.554, "dur": 0.063,
    "args": {
      "Python parent id": 2280, "Python id": 2281, "finished": true, "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902443.868, "dur": 2.053,
    "args": {
      "Python parent id": 2279, "Python id": 2282, "finished": true, "Ev Idx": 5182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.067, "dur": 1.789,
    "args": {
      "Python parent id": 2282, "Python id": 2283, "finished": true, "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.273, "dur": 1.502,
    "args": {
      "Python parent id": 2283, "Python id": 2284, "finished": true, "Ev Idx": 5184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.457, "dur": 1.080,
    "args": {
      "Python parent id": 2284, "Python id": 2285, "finished": true, "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.652, "dur": 0.486,
    "args": {
      "Python parent id": 2285, "Python id": 2286, "finished": true, "Ev Idx": 5186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.752, "dur": 0.045,
    "args": {
      "Python parent id": 2286, "Python id": 2287, "finished": true, "Ev Idx": 5187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902444.974, "dur": 0.118,
    "args": {
      "Python parent id": 2286, "Python id": 2288, "finished": true, "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902446.137, "dur": 0.102,
    "args": {
      "Python parent id": 2279, "Python id": 2289, "finished": true, "Ev Idx": 5189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902446.484, "dur": 0.105,
    "args": {
      "Python parent id": 2277, "Python id": 2290, "finished": true, "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902446.874, "dur": 1.130,
    "args": {
      "Python parent id": 2277, "Python id": 2291, "finished": true, "Ev Idx": 5191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902447.030, "dur": 0.929,
    "args": {
      "Python parent id": 2291, "Python id": 2292, "finished": true, "Ev Idx": 5192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902447.159, "dur": 0.427,
    "args": {
      "Python parent id": 2292, "Python id": 2293, "finished": true, "Ev Idx": 5193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902447.297, "dur": 0.208,
    "args": {
      "Python parent id": 2293, "Python id": 2294, "finished": true, "Ev Idx": 5194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902447.405, "dur": 0.068,
    "args": {
      "Python parent id": 2294, "Python id": 2295, "finished": true, "Ev Idx": 5195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902447.710, "dur": 0.207,
    "args": {
      "Python parent id": 2292, "Python id": 2296, "finished": true, "Ev Idx": 5196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902448.443, "dur": 0.202,
    "args": {
      "Python parent id": 2261, "Python id": 2297, "finished": true, "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902449.318, "dur": 0.973,
    "args": {
      "Python parent id": 2261, "Python id": 2298, "finished": true, "Ev Idx": 5198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902449.725, "dur": 0.515,
    "args": {
      "Python parent id": 2298, "Python id": 2299, "finished": true, "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902450.740, "dur": 2.057,
    "args": {
      "Python parent id": 2260, "Python id": 2300, "finished": true, "Ev Idx": 5200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902453.897, "dur": 95.473,
    "args": {
      "Python parent id": 12, "Python id": 2301, "finished": true, "Ev Idx": 5201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902454.268, "dur": 0.491,
    "args": {
      "Python parent id": 2301, "Python id": 2302, "finished": true, "Ev Idx": 5202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902454.592, "dur": 0.082,
    "args": {
      "Python parent id": 2302, "Python id": 2303, "finished": true, "Ev Idx": 5203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902455.084, "dur": 0.426,
    "args": {
      "Python parent id": 2301, "Python id": 2304, "finished": true, "Ev Idx": 5204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902455.803, "dur": 0.308,
    "args": {
      "Python parent id": 2301, "Python id": 2305, "finished": true, "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902456.352, "dur": 1.701,
    "args": {
      "Python parent id": 2301, "Python id": 2306, "finished": true, "Ev Idx": 5206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902458.714, "dur": 90.531,
    "args": {
      "Python parent id": 2301, "Python id": 2307, "finished": true, "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902459.637, "dur": 89.316,
    "args": {
      "Python parent id": 2307, "Python id": 2308, "finished": true, "Ev Idx": 5208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902459.946, "dur": 88.945,
    "args": {
      "Python parent id": 2308, "Python id": 2309, "finished": true, "Ev Idx": 5209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902466.008, "dur": 81.447,
    "args": {
      "Python parent id": 2309, "Python id": 2310, "finished": true, "Ev Idx": 5210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740902466.347, "dur": 81.002,
    "args": {
      "Python parent id": 2310, "Python id": 2311, "finished": true, "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740902466.973, "dur": 80.306,
    "args": {
      "Python parent id": 2311, "Python id": 2312, "finished": true, "Ev Idx": 5212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902467.544, "dur": 0.421,
    "args": {
      "Python parent id": 2312, "Python id": 2313, "finished": true, "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902467.811, "dur": 0.106,
    "args": {
      "Python parent id": 2313, "Python id": 2314, "finished": true, "Ev Idx": 5214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902468.262, "dur": 0.328,
    "args": {
      "Python parent id": 2312, "Python id": 2315, "finished": true, "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902468.489, "dur": 0.037,
    "args": {
      "Python parent id": 2315, "Python id": 2316, "finished": true, "Ev Idx": 5216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902468.934, "dur": 78.167,
    "args": {
      "Python parent id": 2312, "Python id": 2317, "finished": true, "Ev Idx": 5217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902469.743, "dur": 75.991,
    "args": {
      "Python parent id": 2317, "Python id": 2318, "finished": true, "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740902478.093, "dur": 19.607,
    "args": {
      "Python parent id": 2318, "Python id": 2319, "finished": true, "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740902478.329, "dur": 0.104,
    "args": {
      "Python parent id": 2319, "Python id": 2320, "finished": true, "Ev Idx": 5220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740902478.771, "dur": 18.787,
    "args": {
      "Python parent id": 2319, "Python id": 2321, "finished": true, "Ev Idx": 5221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902478.915, "dur": 8.530,
    "args": {
      "Python parent id": 2321, "Python id": 2322, "finished": true, "Ev Idx": 5222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902505.297, "dur": 7.138,
    "args": {
      "Python parent id": 2318, "Python id": 2323, "finished": true, "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902521.362, "dur": 8.952,
    "args": {
      "Python parent id": 2318, "Python id": 2324, "finished": true, "Ev Idx": 5224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902550.752, "dur": 16.852,
    "args": {
      "Python parent id": 12, "Python id": 2325, "finished": true, "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902551.197, "dur": 13.804,
    "args": {
      "Python parent id": 2325, "Python id": 2326, "finished": true, "Ev Idx": 5226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902551.373, "dur": 0.544,
    "args": {
      "Python parent id": 2326, "Python id": 2327, "finished": true, "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902551.547, "dur": 0.299,
    "args": {
      "Python parent id": 2327, "Python id": 2328, "finished": true, "Ev Idx": 5228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902551.715, "dur": 0.099,
    "args": {
      "Python parent id": 2328, "Python id": 2329, "finished": true, "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902552.333, "dur": 10.097,
    "args": {
      "Python parent id": 2326, "Python id": 2330, "finished": true, "Ev Idx": 5230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902552.446, "dur": 0.071,
    "args": {
      "Python parent id": 2330, "Python id": 2331, "finished": true, "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902552.639, "dur": 0.062,
    "args": {
      "Python parent id": 2330, "Python id": 2332, "finished": true, "Ev Idx": 5232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902552.858, "dur": 0.057,
    "args": {
      "Python parent id": 2330, "Python id": 2333, "finished": true, "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902553.377, "dur": 0.060,
    "args": {
      "Python parent id": 2330, "Python id": 2334, "finished": true, "Ev Idx": 5234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902553.691, "dur": 0.107,
    "args": {
      "Python parent id": 2330, "Python id": 2335, "finished": true, "Ev Idx": 5235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902554.198, "dur": 8.151,
    "args": {
      "Python parent id": 2330, "Python id": 2336, "finished": true, "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902554.298, "dur": 0.071,
    "args": {
      "Python parent id": 2336, "Python id": 2337, "finished": true, "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902554.546, "dur": 0.053,
    "args": {
      "Python parent id": 2336, "Python id": 2338, "finished": true, "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902554.756, "dur": 0.072,
    "args": {
      "Python parent id": 2336, "Python id": 2339, "finished": true, "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902555.236, "dur": 0.029,
    "args": {
      "Python parent id": 2336, "Python id": 2340, "finished": true, "Ev Idx": 5240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902555.587, "dur": 6.647,
    "args": {
      "Python parent id": 2336, "Python id": 2341, "finished": true, "Ev Idx": 5241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902555.917, "dur": 6.225,
    "args": {
      "Python parent id": 2341, "Python id": 2342, "finished": true, "Ev Idx": 5242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902556.085, "dur": 4.104,
    "args": {
      "Python parent id": 2342, "Python id": 2343, "finished": true, "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902556.306, "dur": 3.815,
    "args": {
      "Python parent id": 2343, "Python id": 2344, "finished": true, "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902556.469, "dur": 0.330,
    "args": {
      "Python parent id": 2344, "Python id": 2345, "finished": true, "Ev Idx": 5245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902556.641, "dur": 0.097,
    "args": {
      "Python parent id": 2345, "Python id": 2346, "finished": true, "Ev Idx": 5246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902557.024, "dur": 2.644,
    "args": {
      "Python parent id": 2344, "Python id": 2347, "finished": true, "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902557.262, "dur": 2.287,
    "args": {
      "Python parent id": 2347, "Python id": 2348, "finished": true, "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902557.539, "dur": 1.920,
    "args": {
      "Python parent id": 2348, "Python id": 2349, "finished": true, "Ev Idx": 5249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902557.733, "dur": 1.444,
    "args": {
      "Python parent id": 2349, "Python id": 2350, "finished": true, "Ev Idx": 5250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902557.983, "dur": 0.634,
    "args": {
      "Python parent id": 2350, "Python id": 2351, "finished": true, "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902558.080, "dur": 0.084,
    "args": {
      "Python parent id": 2351, "Python id": 2352, "finished": true, "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902558.367, "dur": 0.192,
    "args": {
      "Python parent id": 2351, "Python id": 2353, "finished": true, "Ev Idx": 5253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902559.890, "dur": 0.163,
    "args": {
      "Python parent id": 2344, "Python id": 2354, "finished": true, "Ev Idx": 5254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902560.385, "dur": 0.105,
    "args": {
      "Python parent id": 2342, "Python id": 2355, "finished": true, "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902560.823, "dur": 1.262,
    "args": {
      "Python parent id": 2342, "Python id": 2356, "finished": true, "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902561.037, "dur": 1.000,
    "args": {
      "Python parent id": 2356, "Python id": 2357, "finished": true, "Ev Idx": 5257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902561.174, "dur": 0.417,
    "args": {
      "Python parent id": 2357, "Python id": 2358, "finished": true, "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902561.315, "dur": 0.211,
    "args": {
      "Python parent id": 2358, "Python id": 2359, "finished": true, "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902561.413, "dur": 0.084,
    "args": {
      "Python parent id": 2359, "Python id": 2360, "finished": true, "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902561.733, "dur": 0.257,
    "args": {
      "Python parent id": 2357, "Python id": 2361, "finished": true, "Ev Idx": 5261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902562.555, "dur": 0.315,
    "args": {
      "Python parent id": 2326, "Python id": 2362, "finished": true, "Ev Idx": 5262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902563.573, "dur": 1.292,
    "args": {
      "Python parent id": 2326, "Python id": 2363, "finished": true, "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902564.154, "dur": 0.663,
    "args": {
      "Python parent id": 2363, "Python id": 2364, "finished": true, "Ev Idx": 5264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902565.291, "dur": 2.206,
    "args": {
      "Python parent id": 2325, "Python id": 2365, "finished": true, "Ev Idx": 5265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902568.307, "dur": 7.349,
    "args": {
      "Python parent id": 12, "Python id": 2366, "finished": true, "Ev Idx": 5266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902576.127, "dur": 0.161,
    "args": {
      "Python parent id": 12, "Python id": 2367, "finished": true, "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902576.955, "dur": 0.226,
    "args": {
      "Python parent id": 12, "Python id": 2368, "finished": true, "Ev Idx": 5268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902577.737, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 2369, "finished": true, "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902578.283, "dur": 10.337,
    "args": {
      "Python parent id": 12, "Python id": 2370, "finished": true, "Ev Idx": 5270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902590.215, "dur": 15.177,
    "args": {
      "Python parent id": 12, "Python id": 2371, "finished": true, "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902590.557, "dur": 12.499,
    "args": {
      "Python parent id": 2371, "Python id": 2372, "finished": true, "Ev Idx": 5272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902590.722, "dur": 0.506,
    "args": {
      "Python parent id": 2372, "Python id": 2373, "finished": true, "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902590.869, "dur": 0.260,
    "args": {
      "Python parent id": 2373, "Python id": 2374, "finished": true, "Ev Idx": 5274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902590.996, "dur": 0.096,
    "args": {
      "Python parent id": 2374, "Python id": 2375, "finished": true, "Ev Idx": 5275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902591.582, "dur": 9.532,
    "args": {
      "Python parent id": 2372, "Python id": 2376, "finished": true, "Ev Idx": 5276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902591.666, "dur": 0.067,
    "args": {
      "Python parent id": 2376, "Python id": 2377, "finished": true, "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902591.855, "dur": 0.060,
    "args": {
      "Python parent id": 2376, "Python id": 2378, "finished": true, "Ev Idx": 5278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902592.125, "dur": 0.052,
    "args": {
      "Python parent id": 2376, "Python id": 2379, "finished": true, "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902592.622, "dur": 0.048,
    "args": {
      "Python parent id": 2376, "Python id": 2380, "finished": true, "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902593.032, "dur": 0.083,
    "args": {
      "Python parent id": 2376, "Python id": 2381, "finished": true, "Ev Idx": 5281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902593.499, "dur": 7.510,
    "args": {
      "Python parent id": 2376, "Python id": 2382, "finished": true, "Ev Idx": 5282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902593.618, "dur": 0.070,
    "args": {
      "Python parent id": 2382, "Python id": 2383, "finished": true, "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902593.900, "dur": 0.058,
    "args": {
      "Python parent id": 2382, "Python id": 2384, "finished": true, "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902594.206, "dur": 0.070,
    "args": {
      "Python parent id": 2382, "Python id": 2385, "finished": true, "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902594.726, "dur": 0.066,
    "args": {
      "Python parent id": 2382, "Python id": 2386, "finished": true, "Ev Idx": 5286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902595.100, "dur": 5.820,
    "args": {
      "Python parent id": 2382, "Python id": 2387, "finished": true, "Ev Idx": 5287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902595.382, "dur": 5.477,
    "args": {
      "Python parent id": 2387, "Python id": 2388, "finished": true, "Ev Idx": 5288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902595.524, "dur": 3.464,
    "args": {
      "Python parent id": 2388, "Python id": 2389, "finished": true, "Ev Idx": 5289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902595.777, "dur": 3.166,
    "args": {
      "Python parent id": 2389, "Python id": 2390, "finished": true, "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902595.931, "dur": 0.317,
    "args": {
      "Python parent id": 2390, "Python id": 2391, "finished": true, "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902596.087, "dur": 0.095,
    "args": {
      "Python parent id": 2391, "Python id": 2392, "finished": true, "Ev Idx": 5292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902596.472, "dur": 2.143,
    "args": {
      "Python parent id": 2390, "Python id": 2393, "finished": true, "Ev Idx": 5293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902596.723, "dur": 1.825,
    "args": {
      "Python parent id": 2393, "Python id": 2394, "finished": true, "Ev Idx": 5294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902596.922, "dur": 1.557,
    "args": {
      "Python parent id": 2394, "Python id": 2395, "finished": true, "Ev Idx": 5295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902597.075, "dur": 1.179,
    "args": {
      "Python parent id": 2395, "Python id": 2396, "finished": true, "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902597.270, "dur": 0.512,
    "args": {
      "Python parent id": 2396, "Python id": 2397, "finished": true, "Ev Idx": 5297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902597.358, "dur": 0.054,
    "args": {
      "Python parent id": 2397, "Python id": 2398, "finished": true, "Ev Idx": 5298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902597.596, "dur": 0.140,
    "args": {
      "Python parent id": 2397, "Python id": 2399, "finished": true, "Ev Idx": 5299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902598.801, "dur": 0.080,
    "args": {
      "Python parent id": 2390, "Python id": 2400, "finished": true, "Ev Idx": 5300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902599.151, "dur": 0.108,
    "args": {
      "Python parent id": 2388, "Python id": 2401, "finished": true, "Ev Idx": 5301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902599.666, "dur": 1.146,
    "args": {
      "Python parent id": 2388, "Python id": 2402, "finished": true, "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902599.799, "dur": 0.970,
    "args": {
      "Python parent id": 2402, "Python id": 2403, "finished": true, "Ev Idx": 5303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902599.941, "dur": 0.397,
    "args": {
      "Python parent id": 2403, "Python id": 2404, "finished": true, "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902600.076, "dur": 0.200,
    "args": {
      "Python parent id": 2404, "Python id": 2405, "finished": true, "Ev Idx": 5305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902600.178, "dur": 0.066,
    "args": {
      "Python parent id": 2405, "Python id": 2406, "finished": true, "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902600.494, "dur": 0.241,
    "args": {
      "Python parent id": 2403, "Python id": 2407, "finished": true, "Ev Idx": 5307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902601.236, "dur": 0.202,
    "args": {
      "Python parent id": 2372, "Python id": 2408, "finished": true, "Ev Idx": 5308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902602.062, "dur": 0.869,
    "args": {
      "Python parent id": 2372, "Python id": 2409, "finished": true, "Ev Idx": 5309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902602.423, "dur": 0.457,
    "args": {
      "Python parent id": 2409, "Python id": 2410, "finished": true, "Ev Idx": 5310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902603.344, "dur": 1.978,
    "args": {
      "Python parent id": 2371, "Python id": 2411, "finished": true, "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902606.316, "dur": 111.437,
    "args": {
      "Python parent id": 12, "Python id": 2412, "finished": true, "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902606.755, "dur": 0.540,
    "args": {
      "Python parent id": 2412, "Python id": 2413, "finished": true, "Ev Idx": 5313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902607.075, "dur": 0.140,
    "args": {
      "Python parent id": 2413, "Python id": 2414, "finished": true, "Ev Idx": 5314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902607.663, "dur": 0.543,
    "args": {
      "Python parent id": 2412, "Python id": 2415, "finished": true, "Ev Idx": 5315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902608.448, "dur": 0.259,
    "args": {
      "Python parent id": 2412, "Python id": 2416, "finished": true, "Ev Idx": 5316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902608.973, "dur": 1.728,
    "args": {
      "Python parent id": 2412, "Python id": 2417, "finished": true, "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902611.339, "dur": 106.348,
    "args": {
      "Python parent id": 2412, "Python id": 2418, "finished": true, "Ev Idx": 5318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902612.272, "dur": 105.140,
    "args": {
      "Python parent id": 2418, "Python id": 2419, "finished": true, "Ev Idx": 5319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902612.572, "dur": 104.715,
    "args": {
      "Python parent id": 2419, "Python id": 2420, "finished": true, "Ev Idx": 5320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902618.095, "dur": 97.169,
    "args": {
      "Python parent id": 2420, "Python id": 2421, "finished": true, "Ev Idx": 5321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740902618.425, "dur": 96.746,
    "args": {
      "Python parent id": 2421, "Python id": 2422, "finished": true, "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740902619.063, "dur": 96.017,
    "args": {
      "Python parent id": 2422, "Python id": 2423, "finished": true, "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902619.539, "dur": 0.454,
    "args": {
      "Python parent id": 2423, "Python id": 2424, "finished": true, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902619.858, "dur": 0.090,
    "args": {
      "Python parent id": 2424, "Python id": 2425, "finished": true, "Ev Idx": 5325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902620.283, "dur": 0.347,
    "args": {
      "Python parent id": 2423, "Python id": 2426, "finished": true, "Ev Idx": 5326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902620.540, "dur": 0.035,
    "args": {
      "Python parent id": 2426, "Python id": 2427, "finished": true, "Ev Idx": 5327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902620.986, "dur": 93.989,
    "args": {
      "Python parent id": 2423, "Python id": 2428, "finished": true, "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902621.724, "dur": 91.467,
    "args": {
      "Python parent id": 2428, "Python id": 2429, "finished": true, "Ev Idx": 5329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740902630.655, "dur": 19.079,
    "args": {
      "Python parent id": 2429, "Python id": 2430, "finished": true, "Ev Idx": 5330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740902630.807, "dur": 0.147,
    "args": {
      "Python parent id": 2430, "Python id": 2431, "finished": true, "Ev Idx": 5331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740902631.309, "dur": 18.225,
    "args": {
      "Python parent id": 2430, "Python id": 2432, "finished": true, "Ev Idx": 5332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902631.472, "dur": 7.960,
    "args": {
      "Python parent id": 2432, "Python id": 2433, "finished": true, "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902657.254, "dur": 6.481,
    "args": {
      "Python parent id": 2429, "Python id": 2434, "finished": true, "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902674.073, "dur": 12.964,
    "args": {
      "Python parent id": 2429, "Python id": 2435, "finished": true, "Ev Idx": 5335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902719.603, "dur": 20.027,
    "args": {
      "Python parent id": 12, "Python id": 2436, "finished": true, "Ev Idx": 5336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902720.158, "dur": 15.867,
    "args": {
      "Python parent id": 2436, "Python id": 2437, "finished": true, "Ev Idx": 5337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902720.364, "dur": 0.724,
    "args": {
      "Python parent id": 2437, "Python id": 2438, "finished": true, "Ev Idx": 5338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902720.560, "dur": 0.411,
    "args": {
      "Python parent id": 2438, "Python id": 2439, "finished": true, "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902720.777, "dur": 0.138,
    "args": {
      "Python parent id": 2439, "Python id": 2440, "finished": true, "Ev Idx": 5340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902721.546, "dur": 11.293,
    "args": {
      "Python parent id": 2437, "Python id": 2441, "finished": true, "Ev Idx": 5341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902721.695, "dur": 0.163,
    "args": {
      "Python parent id": 2441, "Python id": 2442, "finished": true, "Ev Idx": 5342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902722.037, "dur": 0.055,
    "args": {
      "Python parent id": 2441, "Python id": 2443, "finished": true, "Ev Idx": 5343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902722.277, "dur": 0.050,
    "args": {
      "Python parent id": 2441, "Python id": 2444, "finished": true, "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902722.861, "dur": 0.054,
    "args": {
      "Python parent id": 2441, "Python id": 2445, "finished": true, "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902723.224, "dur": 0.079,
    "args": {
      "Python parent id": 2441, "Python id": 2446, "finished": true, "Ev Idx": 5346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902723.803, "dur": 8.979,
    "args": {
      "Python parent id": 2441, "Python id": 2447, "finished": true, "Ev Idx": 5347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902723.925, "dur": 0.066,
    "args": {
      "Python parent id": 2447, "Python id": 2448, "finished": true, "Ev Idx": 5348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902724.200, "dur": 0.054,
    "args": {
      "Python parent id": 2447, "Python id": 2449, "finished": true, "Ev Idx": 5349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902724.439, "dur": 0.069,
    "args": {
      "Python parent id": 2447, "Python id": 2450, "finished": true, "Ev Idx": 5350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902724.940, "dur": 0.057,
    "args": {
      "Python parent id": 2447, "Python id": 2451, "finished": true, "Ev Idx": 5351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902725.335, "dur": 7.341,
    "args": {
      "Python parent id": 2447, "Python id": 2452, "finished": true, "Ev Idx": 5352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902725.656, "dur": 6.916,
    "args": {
      "Python parent id": 2452, "Python id": 2453, "finished": true, "Ev Idx": 5353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902725.816, "dur": 4.567,
    "args": {
      "Python parent id": 2453, "Python id": 2454, "finished": true, "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902726.115, "dur": 4.191,
    "args": {
      "Python parent id": 2454, "Python id": 2455, "finished": true, "Ev Idx": 5355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902726.333, "dur": 0.350,
    "args": {
      "Python parent id": 2455, "Python id": 2456, "finished": true, "Ev Idx": 5356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902726.519, "dur": 0.087,
    "args": {
      "Python parent id": 2456, "Python id": 2457, "finished": true, "Ev Idx": 5357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902726.933, "dur": 2.884,
    "args": {
      "Python parent id": 2455, "Python id": 2458, "finished": true, "Ev Idx": 5358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902727.212, "dur": 2.431,
    "args": {
      "Python parent id": 2458, "Python id": 2459, "finished": true, "Ev Idx": 5359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902727.433, "dur": 2.128,
    "args": {
      "Python parent id": 2459, "Python id": 2460, "finished": true, "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902727.604, "dur": 1.568,
    "args": {
      "Python parent id": 2460, "Python id": 2461, "finished": true, "Ev Idx": 5361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902727.806, "dur": 0.647,
    "args": {
      "Python parent id": 2461, "Python id": 2462, "finished": true, "Ev Idx": 5362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902727.882, "dur": 0.057,
    "args": {
      "Python parent id": 2462, "Python id": 2463, "finished": true, "Ev Idx": 5363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902728.165, "dur": 0.209,
    "args": {
      "Python parent id": 2462, "Python id": 2464, "finished": true, "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902730.029, "dur": 0.185,
    "args": {
      "Python parent id": 2455, "Python id": 2465, "finished": true, "Ev Idx": 5365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902730.603, "dur": 0.121,
    "args": {
      "Python parent id": 2453, "Python id": 2466, "finished": true, "Ev Idx": 5366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902731.105, "dur": 1.406,
    "args": {
      "Python parent id": 2453, "Python id": 2467, "finished": true, "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902731.253, "dur": 1.198,
    "args": {
      "Python parent id": 2467, "Python id": 2468, "finished": true, "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902731.452, "dur": 0.483,
    "args": {
      "Python parent id": 2468, "Python id": 2469, "finished": true, "Ev Idx": 5369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902731.594, "dur": 0.250,
    "args": {
      "Python parent id": 2469, "Python id": 2470, "finished": true, "Ev Idx": 5370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902731.717, "dur": 0.076,
    "args": {
      "Python parent id": 2470, "Python id": 2471, "finished": true, "Ev Idx": 5371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902732.065, "dur": 0.352,
    "args": {
      "Python parent id": 2468, "Python id": 2472, "finished": true, "Ev Idx": 5372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902732.995, "dur": 0.358,
    "args": {
      "Python parent id": 2437, "Python id": 2473, "finished": true, "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902734.173, "dur": 1.668,
    "args": {
      "Python parent id": 2437, "Python id": 2474, "finished": true, "Ev Idx": 5374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902734.880, "dur": 0.851,
    "args": {
      "Python parent id": 2474, "Python id": 2475, "finished": true, "Ev Idx": 5375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902736.470, "dur": 3.022,
    "args": {
      "Python parent id": 2436, "Python id": 2476, "finished": true, "Ev Idx": 5376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902740.281, "dur": 9.367,
    "args": {
      "Python parent id": 12, "Python id": 2477, "finished": true, "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902750.103, "dur": 0.159,
    "args": {
      "Python parent id": 12, "Python id": 2478, "finished": true, "Ev Idx": 5378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902751.056, "dur": 0.280,
    "args": {
      "Python parent id": 12, "Python id": 2479, "finished": true, "Ev Idx": 5379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902751.938, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 2480, "finished": true, "Ev Idx": 5380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902752.427, "dur": 13.395,
    "args": {
      "Python parent id": 12, "Python id": 2481, "finished": true, "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902767.470, "dur": 15.442,
    "args": {
      "Python parent id": 12, "Python id": 2482, "finished": true, "Ev Idx": 5382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902767.857, "dur": 12.753,
    "args": {
      "Python parent id": 2482, "Python id": 2483, "finished": true, "Ev Idx": 5383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902768.058, "dur": 0.567,
    "args": {
      "Python parent id": 2483, "Python id": 2484, "finished": true, "Ev Idx": 5384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902768.196, "dur": 0.305,
    "args": {
      "Python parent id": 2484, "Python id": 2485, "finished": true, "Ev Idx": 5385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902768.352, "dur": 0.115,
    "args": {
      "Python parent id": 2485, "Python id": 2486, "finished": true, "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902768.932, "dur": 9.424,
    "args": {
      "Python parent id": 2483, "Python id": 2487, "finished": true, "Ev Idx": 5387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902769.046, "dur": 0.068,
    "args": {
      "Python parent id": 2487, "Python id": 2488, "finished": true, "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902769.268, "dur": 0.064,
    "args": {
      "Python parent id": 2487, "Python id": 2489, "finished": true, "Ev Idx": 5389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902769.520, "dur": 0.056,
    "args": {
      "Python parent id": 2487, "Python id": 2490, "finished": true, "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902770.106, "dur": 0.082,
    "args": {
      "Python parent id": 2487, "Python id": 2491, "finished": true, "Ev Idx": 5391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902770.460, "dur": 0.073,
    "args": {
      "Python parent id": 2487, "Python id": 2492, "finished": true, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902770.913, "dur": 7.380,
    "args": {
      "Python parent id": 2487, "Python id": 2493, "finished": true, "Ev Idx": 5393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902771.020, "dur": 0.065,
    "args": {
      "Python parent id": 2493, "Python id": 2494, "finished": true, "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902771.273, "dur": 0.063,
    "args": {
      "Python parent id": 2493, "Python id": 2495, "finished": true, "Ev Idx": 5395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902771.526, "dur": 0.075,
    "args": {
      "Python parent id": 2493, "Python id": 2496, "finished": true, "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902771.951, "dur": 0.065,
    "args": {
      "Python parent id": 2493, "Python id": 2497, "finished": true, "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902772.349, "dur": 5.842,
    "args": {
      "Python parent id": 2493, "Python id": 2498, "finished": true, "Ev Idx": 5398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902772.597, "dur": 5.480,
    "args": {
      "Python parent id": 2498, "Python id": 2499, "finished": true, "Ev Idx": 5399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902772.765, "dur": 3.559,
    "args": {
      "Python parent id": 2499, "Python id": 2500, "finished": true, "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902772.991, "dur": 3.259,
    "args": {
      "Python parent id": 2500, "Python id": 2501, "finished": true, "Ev Idx": 5401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902773.137, "dur": 0.341,
    "args": {
      "Python parent id": 2501, "Python id": 2502, "finished": true, "Ev Idx": 5402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902773.297, "dur": 0.098,
    "args": {
      "Python parent id": 2502, "Python id": 2503, "finished": true, "Ev Idx": 5403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902773.680, "dur": 2.221,
    "args": {
      "Python parent id": 2501, "Python id": 2504, "finished": true, "Ev Idx": 5404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902773.917, "dur": 1.919,
    "args": {
      "Python parent id": 2504, "Python id": 2505, "finished": true, "Ev Idx": 5405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902774.100, "dur": 1.670,
    "args": {
      "Python parent id": 2505, "Python id": 2506, "finished": true, "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902774.238, "dur": 1.243,
    "args": {
      "Python parent id": 2506, "Python id": 2507, "finished": true, "Ev Idx": 5407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902774.397, "dur": 0.544,
    "args": {
      "Python parent id": 2507, "Python id": 2508, "finished": true, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902774.485, "dur": 0.047,
    "args": {
      "Python parent id": 2508, "Python id": 2509, "finished": true, "Ev Idx": 5409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902774.733, "dur": 0.145,
    "args": {
      "Python parent id": 2508, "Python id": 2510, "finished": true, "Ev Idx": 5410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902776.079, "dur": 0.077,
    "args": {
      "Python parent id": 2501, "Python id": 2511, "finished": true, "Ev Idx": 5411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902776.496, "dur": 0.076,
    "args": {
      "Python parent id": 2499, "Python id": 2512, "finished": true, "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902776.793, "dur": 1.236,
    "args": {
      "Python parent id": 2499, "Python id": 2513, "finished": true, "Ev Idx": 5413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902776.947, "dur": 1.042,
    "args": {
      "Python parent id": 2513, "Python id": 2514, "finished": true, "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902777.121, "dur": 0.445,
    "args": {
      "Python parent id": 2514, "Python id": 2515, "finished": true, "Ev Idx": 5415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902777.272, "dur": 0.210,
    "args": {
      "Python parent id": 2515, "Python id": 2516, "finished": true, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902777.372, "dur": 0.076,
    "args": {
      "Python parent id": 2516, "Python id": 2517, "finished": true, "Ev Idx": 5417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902777.699, "dur": 0.259,
    "args": {
      "Python parent id": 2514, "Python id": 2518, "finished": true, "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902778.490, "dur": 0.251,
    "args": {
      "Python parent id": 2483, "Python id": 2519, "finished": true, "Ev Idx": 5419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902779.385, "dur": 1.078,
    "args": {
      "Python parent id": 2483, "Python id": 2520, "finished": true, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902779.866, "dur": 0.544,
    "args": {
      "Python parent id": 2520, "Python id": 2521, "finished": true, "Ev Idx": 5421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902780.929, "dur": 1.880,
    "args": {
      "Python parent id": 2482, "Python id": 2522, "finished": true, "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902783.838, "dur": 117.754,
    "args": {
      "Python parent id": 12, "Python id": 2523, "finished": true, "Ev Idx": 5423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902784.313, "dur": 0.504,
    "args": {
      "Python parent id": 2523, "Python id": 2524, "finished": true, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902784.636, "dur": 0.121,
    "args": {
      "Python parent id": 2524, "Python id": 2525, "finished": true, "Ev Idx": 5425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902785.180, "dur": 0.455,
    "args": {
      "Python parent id": 2523, "Python id": 2526, "finished": true, "Ev Idx": 5426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902785.855, "dur": 0.263,
    "args": {
      "Python parent id": 2523, "Python id": 2527, "finished": true, "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902786.373, "dur": 1.936,
    "args": {
      "Python parent id": 2523, "Python id": 2528, "finished": true, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902789.013, "dur": 112.479,
    "args": {
      "Python parent id": 2523, "Python id": 2529, "finished": true, "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902789.981, "dur": 111.244,
    "args": {
      "Python parent id": 2529, "Python id": 2530, "finished": true, "Ev Idx": 5430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902790.342, "dur": 110.820,
    "args": {
      "Python parent id": 2530, "Python id": 2531, "finished": true, "Ev Idx": 5431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740902798.780, "dur": 100.755,
    "args": {
      "Python parent id": 2531, "Python id": 2532, "finished": true, "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740902799.117, "dur": 100.270,
    "args": {
      "Python parent id": 2532, "Python id": 2533, "finished": true, "Ev Idx": 5433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740902799.733, "dur": 99.578,
    "args": {
      "Python parent id": 2533, "Python id": 2534, "finished": true, "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902800.220, "dur": 0.452,
    "args": {
      "Python parent id": 2534, "Python id": 2535, "finished": true, "Ev Idx": 5435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902800.519, "dur": 0.078,
    "args": {
      "Python parent id": 2535, "Python id": 2536, "finished": true, "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902801.007, "dur": 0.359,
    "args": {
      "Python parent id": 2534, "Python id": 2537, "finished": true, "Ev Idx": 5437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902801.264, "dur": 0.060,
    "args": {
      "Python parent id": 2537, "Python id": 2538, "finished": true, "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902801.716, "dur": 97.491,
    "args": {
      "Python parent id": 2534, "Python id": 2539, "finished": true, "Ev Idx": 5439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740902802.495, "dur": 95.384,
    "args": {
      "Python parent id": 2539, "Python id": 2540, "finished": true, "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740902811.748, "dur": 32.143,
    "args": {
      "Python parent id": 2540, "Python id": 2541, "finished": true, "Ev Idx": 5441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740902811.922, "dur": 0.134,
    "args": {
      "Python parent id": 2541, "Python id": 2542, "finished": true, "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740902812.431, "dur": 31.271,
    "args": {
      "Python parent id": 2541, "Python id": 2543, "finished": true, "Ev Idx": 5443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902812.637, "dur": 19.097,
    "args": {
      "Python parent id": 2543, "Python id": 2544, "finished": true, "Ev Idx": 5444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902852.818, "dur": 8.184,
    "args": {
      "Python parent id": 2540, "Python id": 2545, "finished": true, "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740902871.421, "dur": 10.089,
    "args": {
      "Python parent id": 2540, "Python id": 2546, "finished": true, "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902902.982, "dur": 21.064,
    "args": {
      "Python parent id": 12, "Python id": 2547, "finished": true, "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902903.432, "dur": 16.939,
    "args": {
      "Python parent id": 2547, "Python id": 2548, "finished": true, "Ev Idx": 5448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902903.599, "dur": 0.558,
    "args": {
      "Python parent id": 2548, "Python id": 2549, "finished": true, "Ev Idx": 5449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902903.808, "dur": 0.240,
    "args": {
      "Python parent id": 2549, "Python id": 2550, "finished": true, "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902903.937, "dur": 0.076,
    "args": {
      "Python parent id": 2550, "Python id": 2551, "finished": true, "Ev Idx": 5451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902904.681, "dur": 11.445,
    "args": {
      "Python parent id": 2548, "Python id": 2552, "finished": true, "Ev Idx": 5452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902904.791, "dur": 0.075,
    "args": {
      "Python parent id": 2552, "Python id": 2553, "finished": true, "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902905.001, "dur": 0.062,
    "args": {
      "Python parent id": 2552, "Python id": 2554, "finished": true, "Ev Idx": 5454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902905.284, "dur": 0.052,
    "args": {
      "Python parent id": 2552, "Python id": 2555, "finished": true, "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902905.898, "dur": 0.060,
    "args": {
      "Python parent id": 2552, "Python id": 2556, "finished": true, "Ev Idx": 5456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902906.307, "dur": 0.071,
    "args": {
      "Python parent id": 2552, "Python id": 2557, "finished": true, "Ev Idx": 5457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902906.868, "dur": 9.188,
    "args": {
      "Python parent id": 2552, "Python id": 2558, "finished": true, "Ev Idx": 5458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902906.958, "dur": 0.064,
    "args": {
      "Python parent id": 2558, "Python id": 2559, "finished": true, "Ev Idx": 5459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902907.183, "dur": 0.068,
    "args": {
      "Python parent id": 2558, "Python id": 2560, "finished": true, "Ev Idx": 5460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902907.455, "dur": 0.072,
    "args": {
      "Python parent id": 2558, "Python id": 2561, "finished": true, "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902907.943, "dur": 0.058,
    "args": {
      "Python parent id": 2558, "Python id": 2562, "finished": true, "Ev Idx": 5462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902908.409, "dur": 7.487,
    "args": {
      "Python parent id": 2558, "Python id": 2563, "finished": true, "Ev Idx": 5463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902908.736, "dur": 7.048,
    "args": {
      "Python parent id": 2563, "Python id": 2564, "finished": true, "Ev Idx": 5464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902908.891, "dur": 4.177,
    "args": {
      "Python parent id": 2564, "Python id": 2565, "finished": true, "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902909.113, "dur": 3.875,
    "args": {
      "Python parent id": 2565, "Python id": 2566, "finished": true, "Ev Idx": 5466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902909.330, "dur": 0.393,
    "args": {
      "Python parent id": 2566, "Python id": 2567, "finished": true, "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902909.556, "dur": 0.065,
    "args": {
      "Python parent id": 2567, "Python id": 2568, "finished": true, "Ev Idx": 5468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902909.920, "dur": 2.648,
    "args": {
      "Python parent id": 2566, "Python id": 2569, "finished": true, "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902910.216, "dur": 2.223,
    "args": {
      "Python parent id": 2569, "Python id": 2570, "finished": true, "Ev Idx": 5470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902910.424, "dur": 1.923,
    "args": {
      "Python parent id": 2570, "Python id": 2571, "finished": true, "Ev Idx": 5471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902910.588, "dur": 1.475,
    "args": {
      "Python parent id": 2571, "Python id": 2572, "finished": true, "Ev Idx": 5472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902910.811, "dur": 0.656,
    "args": {
      "Python parent id": 2572, "Python id": 2573, "finished": true, "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902910.893, "dur": 0.052,
    "args": {
      "Python parent id": 2573, "Python id": 2574, "finished": true, "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902911.179, "dur": 0.208,
    "args": {
      "Python parent id": 2573, "Python id": 2575, "finished": true, "Ev Idx": 5475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902912.787, "dur": 0.131,
    "args": {
      "Python parent id": 2566, "Python id": 2576, "finished": true, "Ev Idx": 5476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902913.268, "dur": 0.096,
    "args": {
      "Python parent id": 2564, "Python id": 2577, "finished": true, "Ev Idx": 5477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902913.678, "dur": 2.058,
    "args": {
      "Python parent id": 2564, "Python id": 2578, "finished": true, "Ev Idx": 5478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902914.523, "dur": 1.148,
    "args": {
      "Python parent id": 2578, "Python id": 2579, "finished": true, "Ev Idx": 5479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902914.688, "dur": 0.415,
    "args": {
      "Python parent id": 2579, "Python id": 2580, "finished": true, "Ev Idx": 5480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902914.826, "dur": 0.193,
    "args": {
      "Python parent id": 2580, "Python id": 2581, "finished": true, "Ev Idx": 5481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902914.930, "dur": 0.054,
    "args": {
      "Python parent id": 2581, "Python id": 2582, "finished": true, "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902915.250, "dur": 0.388,
    "args": {
      "Python parent id": 2579, "Python id": 2583, "finished": true, "Ev Idx": 5483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902916.268, "dur": 0.269,
    "args": {
      "Python parent id": 2548, "Python id": 2584, "finished": true, "Ev Idx": 5484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902918.713, "dur": 1.493,
    "args": {
      "Python parent id": 2548, "Python id": 2585, "finished": true, "Ev Idx": 5485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902919.318, "dur": 0.764,
    "args": {
      "Python parent id": 2585, "Python id": 2586, "finished": true, "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902920.772, "dur": 3.183,
    "args": {
      "Python parent id": 2547, "Python id": 2587, "finished": true, "Ev Idx": 5487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902924.661, "dur": 7.935,
    "args": {
      "Python parent id": 12, "Python id": 2588, "finished": true, "Ev Idx": 5488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740902933.136, "dur": 0.122,
    "args": {
      "Python parent id": 12, "Python id": 2589, "finished": true, "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902933.955, "dur": 0.205,
    "args": {
      "Python parent id": 12, "Python id": 2590, "finished": true, "Ev Idx": 5490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902934.753, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 2591, "finished": true, "Ev Idx": 5491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740902935.262, "dur": 35.335,
    "args": {
      "Python parent id": 12, "Python id": 2592, "finished": true, "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740902972.075, "dur": 15.627,
    "args": {
      "Python parent id": 12, "Python id": 2593, "finished": true, "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740902972.451, "dur": 12.220,
    "args": {
      "Python parent id": 2593, "Python id": 2594, "finished": true, "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902972.603, "dur": 0.607,
    "args": {
      "Python parent id": 2594, "Python id": 2595, "finished": true, "Ev Idx": 5495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902972.801, "dur": 0.304,
    "args": {
      "Python parent id": 2595, "Python id": 2596, "finished": true, "Ev Idx": 5496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902972.934, "dur": 0.115,
    "args": {
      "Python parent id": 2596, "Python id": 2597, "finished": true, "Ev Idx": 5497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902973.575, "dur": 9.044,
    "args": {
      "Python parent id": 2594, "Python id": 2598, "finished": true, "Ev Idx": 5498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902973.662, "dur": 0.070,
    "args": {
      "Python parent id": 2598, "Python id": 2599, "finished": true, "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902973.874, "dur": 0.061,
    "args": {
      "Python parent id": 2598, "Python id": 2600, "finished": true, "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902974.106, "dur": 0.050,
    "args": {
      "Python parent id": 2598, "Python id": 2601, "finished": true, "Ev Idx": 5501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902974.636, "dur": 0.040,
    "args": {
      "Python parent id": 2598, "Python id": 2602, "finished": true, "Ev Idx": 5502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902974.969, "dur": 0.069,
    "args": {
      "Python parent id": 2598, "Python id": 2603, "finished": true, "Ev Idx": 5503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902975.361, "dur": 7.188,
    "args": {
      "Python parent id": 2598, "Python id": 2604, "finished": true, "Ev Idx": 5504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902975.440, "dur": 0.063,
    "args": {
      "Python parent id": 2604, "Python id": 2605, "finished": true, "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902975.692, "dur": 0.051,
    "args": {
      "Python parent id": 2604, "Python id": 2606, "finished": true, "Ev Idx": 5506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902975.944, "dur": 0.066,
    "args": {
      "Python parent id": 2604, "Python id": 2607, "finished": true, "Ev Idx": 5507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740902976.329, "dur": 0.039,
    "args": {
      "Python parent id": 2604, "Python id": 2608, "finished": true, "Ev Idx": 5508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740902976.654, "dur": 5.791,
    "args": {
      "Python parent id": 2604, "Python id": 2609, "finished": true, "Ev Idx": 5509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740902976.927, "dur": 5.404,
    "args": {
      "Python parent id": 2609, "Python id": 2610, "finished": true, "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740902977.050, "dur": 3.496,
    "args": {
      "Python parent id": 2610, "Python id": 2611, "finished": true, "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740902977.232, "dur": 3.239,
    "args": {
      "Python parent id": 2611, "Python id": 2612, "finished": true, "Ev Idx": 5512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740902977.442, "dur": 0.301,
    "args": {
      "Python parent id": 2612, "Python id": 2613, "finished": true, "Ev Idx": 5513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740902977.625, "dur": 0.061,
    "args": {
      "Python parent id": 2613, "Python id": 2614, "finished": true, "Ev Idx": 5514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740902977.919, "dur": 2.237,
    "args": {
      "Python parent id": 2612, "Python id": 2615, "finished": true, "Ev Idx": 5515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740902978.169, "dur": 1.898,
    "args": {
      "Python parent id": 2615, "Python id": 2616, "finished": true, "Ev Idx": 5516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740902978.377, "dur": 1.621,
    "args": {
      "Python parent id": 2616, "Python id": 2617, "finished": true, "Ev Idx": 5517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740902978.523, "dur": 1.270,
    "args": {
      "Python parent id": 2617, "Python id": 2618, "finished": true, "Ev Idx": 5518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740902978.713, "dur": 0.531,
    "args": {
      "Python parent id": 2618, "Python id": 2619, "finished": true, "Ev Idx": 5519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740902978.797, "dur": 0.049,
    "args": {
      "Python parent id": 2619, "Python id": 2620, "finished": true, "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740902979.055, "dur": 0.132,
    "args": {
      "Python parent id": 2619, "Python id": 2621, "finished": true, "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740902980.348, "dur": 0.070,
    "args": {
      "Python parent id": 2612, "Python id": 2622, "finished": true, "Ev Idx": 5522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740902980.680, "dur": 0.082,
    "args": {
      "Python parent id": 2610, "Python id": 2623, "finished": true, "Ev Idx": 5523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.012, "dur": 1.273,
    "args": {
      "Python parent id": 2610, "Python id": 2624, "finished": true, "Ev Idx": 5524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.168, "dur": 1.070,
    "args": {
      "Python parent id": 2624, "Python id": 2625, "finished": true, "Ev Idx": 5525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.315, "dur": 0.409,
    "args": {
      "Python parent id": 2625, "Python id": 2626, "finished": true, "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.455, "dur": 0.198,
    "args": {
      "Python parent id": 2626, "Python id": 2627, "finished": true, "Ev Idx": 5527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.559, "dur": 0.063,
    "args": {
      "Python parent id": 2627, "Python id": 2628, "finished": true, "Ev Idx": 5528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740902981.856, "dur": 0.347,
    "args": {
      "Python parent id": 2625, "Python id": 2629, "finished": true, "Ev Idx": 5529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740902982.747, "dur": 0.181,
    "args": {
      "Python parent id": 2594, "Python id": 2630, "finished": true, "Ev Idx": 5530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740902983.569, "dur": 0.972,
    "args": {
      "Python parent id": 2594, "Python id": 2631, "finished": true, "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740902983.971, "dur": 0.520,
    "args": {
      "Python parent id": 2631, "Python id": 2632, "finished": true, "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740902984.943, "dur": 2.689,
    "args": {
      "Python parent id": 2593, "Python id": 2633, "finished": true, "Ev Idx": 5533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740902988.553, "dur": 108.811,
    "args": {
      "Python parent id": 12, "Python id": 2634, "finished": true, "Ev Idx": 5534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740902989.031, "dur": 0.506,
    "args": {
      "Python parent id": 2634, "Python id": 2635, "finished": true, "Ev Idx": 5535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740902989.338, "dur": 0.121,
    "args": {
      "Python parent id": 2635, "Python id": 2636, "finished": true, "Ev Idx": 5536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902989.870, "dur": 0.490,
    "args": {
      "Python parent id": 2634, "Python id": 2637, "finished": true, "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740902990.631, "dur": 0.252,
    "args": {
      "Python parent id": 2634, "Python id": 2638, "finished": true, "Ev Idx": 5538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740902991.298, "dur": 1.895,
    "args": {
      "Python parent id": 2634, "Python id": 2639, "finished": true, "Ev Idx": 5539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740902993.885, "dur": 103.345,
    "args": {
      "Python parent id": 2634, "Python id": 2640, "finished": true, "Ev Idx": 5540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740902994.982, "dur": 102.058,
    "args": {
      "Python parent id": 2640, "Python id": 2641, "finished": true, "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740902995.367, "dur": 101.612,
    "args": {
      "Python parent id": 2641, "Python id": 2642, "finished": true, "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903004.308, "dur": 91.052,
    "args": {
      "Python parent id": 2642, "Python id": 2643, "finished": true, "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903004.677, "dur": 90.620,
    "args": {
      "Python parent id": 2643, "Python id": 2644, "finished": true, "Ev Idx": 5544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903005.350, "dur": 89.687,
    "args": {
      "Python parent id": 2644, "Python id": 2645, "finished": true, "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903005.915, "dur": 0.362,
    "args": {
      "Python parent id": 2645, "Python id": 2646, "finished": true, "Ev Idx": 5546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903006.139, "dur": 0.060,
    "args": {
      "Python parent id": 2646, "Python id": 2647, "finished": true, "Ev Idx": 5547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903006.641, "dur": 0.430,
    "args": {
      "Python parent id": 2645, "Python id": 2648, "finished": true, "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903006.931, "dur": 0.086,
    "args": {
      "Python parent id": 2648, "Python id": 2649, "finished": true, "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903007.502, "dur": 87.412,
    "args": {
      "Python parent id": 2645, "Python id": 2650, "finished": true, "Ev Idx": 5550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903008.233, "dur": 85.198,
    "args": {
      "Python parent id": 2650, "Python id": 2651, "finished": true, "Ev Idx": 5551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903019.735, "dur": 21.703,
    "args": {
      "Python parent id": 2651, "Python id": 2652, "finished": true, "Ev Idx": 5552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903019.994, "dur": 0.092,
    "args": {
      "Python parent id": 2652, "Python id": 2653, "finished": true, "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903020.441, "dur": 20.845,
    "args": {
      "Python parent id": 2652, "Python id": 2654, "finished": true, "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903020.651, "dur": 8.965,
    "args": {
      "Python parent id": 2654, "Python id": 2655, "finished": true, "Ev Idx": 5555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903049.943, "dur": 7.741,
    "args": {
      "Python parent id": 2651, "Python id": 2656, "finished": true, "Ev Idx": 5556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903067.485, "dur": 9.639,
    "args": {
      "Python parent id": 2651, "Python id": 2657, "finished": true, "Ev Idx": 5557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903098.684, "dur": 19.382,
    "args": {
      "Python parent id": 12, "Python id": 2658, "finished": true, "Ev Idx": 5558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903099.108, "dur": 13.756,
    "args": {
      "Python parent id": 2658, "Python id": 2659, "finished": true, "Ev Idx": 5559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903099.284, "dur": 0.555,
    "args": {
      "Python parent id": 2659, "Python id": 2660, "finished": true, "Ev Idx": 5560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903099.479, "dur": 0.265,
    "args": {
      "Python parent id": 2660, "Python id": 2661, "finished": true, "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903099.608, "dur": 0.108,
    "args": {
      "Python parent id": 2661, "Python id": 2662, "finished": true, "Ev Idx": 5562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903100.213, "dur": 10.176,
    "args": {
      "Python parent id": 2659, "Python id": 2663, "finished": true, "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903100.374, "dur": 0.102,
    "args": {
      "Python parent id": 2663, "Python id": 2664, "finished": true, "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903100.616, "dur": 0.058,
    "args": {
      "Python parent id": 2663, "Python id": 2665, "finished": true, "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903100.912, "dur": 0.055,
    "args": {
      "Python parent id": 2663, "Python id": 2666, "finished": true, "Ev Idx": 5566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903101.482, "dur": 0.050,
    "args": {
      "Python parent id": 2663, "Python id": 2667, "finished": true, "Ev Idx": 5567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903101.886, "dur": 0.087,
    "args": {
      "Python parent id": 2663, "Python id": 2668, "finished": true, "Ev Idx": 5568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903102.402, "dur": 7.926,
    "args": {
      "Python parent id": 2663, "Python id": 2669, "finished": true, "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903102.501, "dur": 0.068,
    "args": {
      "Python parent id": 2669, "Python id": 2670, "finished": true, "Ev Idx": 5570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903102.742, "dur": 0.060,
    "args": {
      "Python parent id": 2669, "Python id": 2671, "finished": true, "Ev Idx": 5571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903103.046, "dur": 0.066,
    "args": {
      "Python parent id": 2669, "Python id": 2672, "finished": true, "Ev Idx": 5572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903103.478, "dur": 0.039,
    "args": {
      "Python parent id": 2669, "Python id": 2673, "finished": true, "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903103.845, "dur": 6.380,
    "args": {
      "Python parent id": 2669, "Python id": 2674, "finished": true, "Ev Idx": 5574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903104.133, "dur": 6.018,
    "args": {
      "Python parent id": 2674, "Python id": 2675, "finished": true, "Ev Idx": 5575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903104.328, "dur": 3.922,
    "args": {
      "Python parent id": 2675, "Python id": 2676, "finished": true, "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903104.569, "dur": 3.600,
    "args": {
      "Python parent id": 2676, "Python id": 2677, "finished": true, "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903104.720, "dur": 0.393,
    "args": {
      "Python parent id": 2677, "Python id": 2678, "finished": true, "Ev Idx": 5578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903104.937, "dur": 0.056,
    "args": {
      "Python parent id": 2678, "Python id": 2679, "finished": true, "Ev Idx": 5579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903105.299, "dur": 2.497,
    "args": {
      "Python parent id": 2677, "Python id": 2680, "finished": true, "Ev Idx": 5580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903105.608, "dur": 2.088,
    "args": {
      "Python parent id": 2680, "Python id": 2681, "finished": true, "Ev Idx": 5581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903105.870, "dur": 1.759,
    "args": {
      "Python parent id": 2681, "Python id": 2682, "finished": true, "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903106.071, "dur": 1.342,
    "args": {
      "Python parent id": 2682, "Python id": 2683, "finished": true, "Ev Idx": 5583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903106.337, "dur": 0.575,
    "args": {
      "Python parent id": 2683, "Python id": 2684, "finished": true, "Ev Idx": 5584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903106.448, "dur": 0.044,
    "args": {
      "Python parent id": 2684, "Python id": 2685, "finished": true, "Ev Idx": 5585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903106.700, "dur": 0.156,
    "args": {
      "Python parent id": 2684, "Python id": 2686, "finished": true, "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903108.020, "dur": 0.117,
    "args": {
      "Python parent id": 2677, "Python id": 2687, "finished": true, "Ev Idx": 5587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903108.429, "dur": 0.072,
    "args": {
      "Python parent id": 2675, "Python id": 2688, "finished": true, "Ev Idx": 5588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903108.838, "dur": 1.271,
    "args": {
      "Python parent id": 2675, "Python id": 2689, "finished": true, "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903109.019, "dur": 1.043,
    "args": {
      "Python parent id": 2689, "Python id": 2690, "finished": true, "Ev Idx": 5590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903109.174, "dur": 0.431,
    "args": {
      "Python parent id": 2690, "Python id": 2691, "finished": true, "Ev Idx": 5591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903109.366, "dur": 0.175,
    "args": {
      "Python parent id": 2691, "Python id": 2692, "finished": true, "Ev Idx": 5592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903109.451, "dur": 0.055,
    "args": {
      "Python parent id": 2692, "Python id": 2693, "finished": true, "Ev Idx": 5593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903109.746, "dur": 0.286,
    "args": {
      "Python parent id": 2690, "Python id": 2694, "finished": true, "Ev Idx": 5594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903110.529, "dur": 0.305,
    "args": {
      "Python parent id": 2659, "Python id": 2695, "finished": true, "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903111.503, "dur": 1.228,
    "args": {
      "Python parent id": 2659, "Python id": 2696, "finished": true, "Ev Idx": 5596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903112.060, "dur": 0.620,
    "args": {
      "Python parent id": 2696, "Python id": 2697, "finished": true, "Ev Idx": 5597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903113.177, "dur": 4.818,
    "args": {
      "Python parent id": 2658, "Python id": 2698, "finished": true, "Ev Idx": 5598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903118.648, "dur": 8.297,
    "args": {
      "Python parent id": 12, "Python id": 2699, "finished": true, "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903127.488, "dur": 0.128,
    "args": {
      "Python parent id": 12, "Python id": 2700, "finished": true, "Ev Idx": 5600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903128.376, "dur": 0.232,
    "args": {
      "Python parent id": 12, "Python id": 2701, "finished": true, "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903129.178, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 2702, "finished": true, "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903129.654, "dur": 10.883,
    "args": {
      "Python parent id": 12, "Python id": 2703, "finished": true, "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903142.003, "dur": 14.375,
    "args": {
      "Python parent id": 12, "Python id": 2704, "finished": true, "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903142.289, "dur": 11.805,
    "args": {
      "Python parent id": 2704, "Python id": 2705, "finished": true, "Ev Idx": 5605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903142.476, "dur": 0.506,
    "args": {
      "Python parent id": 2705, "Python id": 2706, "finished": true, "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903142.625, "dur": 0.236,
    "args": {
      "Python parent id": 2706, "Python id": 2707, "finished": true, "Ev Idx": 5607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903142.728, "dur": 0.098,
    "args": {
      "Python parent id": 2707, "Python id": 2708, "finished": true, "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903143.250, "dur": 8.841,
    "args": {
      "Python parent id": 2705, "Python id": 2709, "finished": true, "Ev Idx": 5609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903143.351, "dur": 0.088,
    "args": {
      "Python parent id": 2709, "Python id": 2710, "finished": true, "Ev Idx": 5610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903143.563, "dur": 0.082,
    "args": {
      "Python parent id": 2709, "Python id": 2711, "finished": true, "Ev Idx": 5611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903143.822, "dur": 0.071,
    "args": {
      "Python parent id": 2709, "Python id": 2712, "finished": true, "Ev Idx": 5612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903144.370, "dur": 0.054,
    "args": {
      "Python parent id": 2709, "Python id": 2713, "finished": true, "Ev Idx": 5613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903144.710, "dur": 0.089,
    "args": {
      "Python parent id": 2709, "Python id": 2714, "finished": true, "Ev Idx": 5614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903145.145, "dur": 6.882,
    "args": {
      "Python parent id": 2709, "Python id": 2715, "finished": true, "Ev Idx": 5615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903145.223, "dur": 0.060,
    "args": {
      "Python parent id": 2715, "Python id": 2716, "finished": true, "Ev Idx": 5616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903145.425, "dur": 0.056,
    "args": {
      "Python parent id": 2715, "Python id": 2717, "finished": true, "Ev Idx": 5617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903145.669, "dur": 0.062,
    "args": {
      "Python parent id": 2715, "Python id": 2718, "finished": true, "Ev Idx": 5618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903146.071, "dur": 0.066,
    "args": {
      "Python parent id": 2715, "Python id": 2719, "finished": true, "Ev Idx": 5619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903146.457, "dur": 5.485,
    "args": {
      "Python parent id": 2715, "Python id": 2720, "finished": true, "Ev Idx": 5620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903146.707, "dur": 5.176,
    "args": {
      "Python parent id": 2720, "Python id": 2721, "finished": true, "Ev Idx": 5621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903146.881, "dur": 3.278,
    "args": {
      "Python parent id": 2721, "Python id": 2722, "finished": true, "Ev Idx": 5622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903147.073, "dur": 3.042,
    "args": {
      "Python parent id": 2722, "Python id": 2723, "finished": true, "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903147.231, "dur": 0.336,
    "args": {
      "Python parent id": 2723, "Python id": 2724, "finished": true, "Ev Idx": 5624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903147.397, "dur": 0.070,
    "args": {
      "Python parent id": 2724, "Python id": 2725, "finished": true, "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903147.717, "dur": 2.057,
    "args": {
      "Python parent id": 2723, "Python id": 2726, "finished": true, "Ev Idx": 5626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903147.970, "dur": 1.743,
    "args": {
      "Python parent id": 2726, "Python id": 2727, "finished": true, "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903148.185, "dur": 1.487,
    "args": {
      "Python parent id": 2727, "Python id": 2728, "finished": true, "Ev Idx": 5628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903148.343, "dur": 1.079,
    "args": {
      "Python parent id": 2728, "Python id": 2729, "finished": true, "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903148.545, "dur": 0.482,
    "args": {
      "Python parent id": 2729, "Python id": 2730, "finished": true, "Ev Idx": 5630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903148.641, "dur": 0.052,
    "args": {
      "Python parent id": 2730, "Python id": 2731, "finished": true, "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903148.847, "dur": 0.119,
    "args": {
      "Python parent id": 2730, "Python id": 2732, "finished": true, "Ev Idx": 5632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903149.996, "dur": 0.070,
    "args": {
      "Python parent id": 2723, "Python id": 2733, "finished": true, "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903150.325, "dur": 0.082,
    "args": {
      "Python parent id": 2721, "Python id": 2734, "finished": true, "Ev Idx": 5634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903150.673, "dur": 1.162,
    "args": {
      "Python parent id": 2721, "Python id": 2735, "finished": true, "Ev Idx": 5635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903150.841, "dur": 0.950,
    "args": {
      "Python parent id": 2735, "Python id": 2736, "finished": true, "Ev Idx": 5636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903150.988, "dur": 0.415,
    "args": {
      "Python parent id": 2736, "Python id": 2737, "finished": true, "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903151.142, "dur": 0.196,
    "args": {
      "Python parent id": 2737, "Python id": 2738, "finished": true, "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903151.241, "dur": 0.067,
    "args": {
      "Python parent id": 2738, "Python id": 2739, "finished": true, "Ev Idx": 5639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903151.541, "dur": 0.220,
    "args": {
      "Python parent id": 2736, "Python id": 2740, "finished": true, "Ev Idx": 5640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903152.240, "dur": 0.159,
    "args": {
      "Python parent id": 2705, "Python id": 2741, "finished": true, "Ev Idx": 5641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903153.023, "dur": 0.913,
    "args": {
      "Python parent id": 2705, "Python id": 2742, "finished": true, "Ev Idx": 5642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903153.413, "dur": 0.460,
    "args": {
      "Python parent id": 2742, "Python id": 2743, "finished": true, "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903154.384, "dur": 1.926,
    "args": {
      "Python parent id": 2704, "Python id": 2744, "finished": true, "Ev Idx": 5644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903157.261, "dur": 100.902,
    "args": {
      "Python parent id": 12, "Python id": 2745, "finished": true, "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903157.780, "dur": 0.778,
    "args": {
      "Python parent id": 2745, "Python id": 2746, "finished": true, "Ev Idx": 5646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903158.137, "dur": 0.335,
    "args": {
      "Python parent id": 2746, "Python id": 2747, "finished": true, "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903159.169, "dur": 0.477,
    "args": {
      "Python parent id": 2745, "Python id": 2748, "finished": true, "Ev Idx": 5648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903159.931, "dur": 0.280,
    "args": {
      "Python parent id": 2745, "Python id": 2749, "finished": true, "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903160.630, "dur": 1.604,
    "args": {
      "Python parent id": 2745, "Python id": 2750, "finished": true, "Ev Idx": 5650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903162.884, "dur": 95.169,
    "args": {
      "Python parent id": 2745, "Python id": 2751, "finished": true, "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903164.281, "dur": 93.565,
    "args": {
      "Python parent id": 2751, "Python id": 2752, "finished": true, "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903164.562, "dur": 93.228,
    "args": {
      "Python parent id": 2752, "Python id": 2753, "finished": true, "Ev Idx": 5653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903170.553, "dur": 85.754,
    "args": {
      "Python parent id": 2753, "Python id": 2754, "finished": true, "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903170.883, "dur": 85.368,
    "args": {
      "Python parent id": 2754, "Python id": 2755, "finished": true, "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903171.493, "dur": 84.559,
    "args": {
      "Python parent id": 2755, "Python id": 2756, "finished": true, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903172.243, "dur": 0.608,
    "args": {
      "Python parent id": 2756, "Python id": 2757, "finished": true, "Ev Idx": 5657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903172.524, "dur": 0.074,
    "args": {
      "Python parent id": 2757, "Python id": 2758, "finished": true, "Ev Idx": 5658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903173.273, "dur": 0.321,
    "args": {
      "Python parent id": 2756, "Python id": 2759, "finished": true, "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903173.471, "dur": 0.076,
    "args": {
      "Python parent id": 2759, "Python id": 2760, "finished": true, "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903174.043, "dur": 81.885,
    "args": {
      "Python parent id": 2756, "Python id": 2761, "finished": true, "Ev Idx": 5661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903175.277, "dur": 79.161,
    "args": {
      "Python parent id": 2761, "Python id": 2762, "finished": true, "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903184.492, "dur": 20.031,
    "args": {
      "Python parent id": 2762, "Python id": 2763, "finished": true, "Ev Idx": 5663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903184.676, "dur": 0.086,
    "args": {
      "Python parent id": 2763, "Python id": 2764, "finished": true, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903185.167, "dur": 19.235,
    "args": {
      "Python parent id": 2763, "Python id": 2765, "finished": true, "Ev Idx": 5665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903185.363, "dur": 7.821,
    "args": {
      "Python parent id": 2765, "Python id": 2766, "finished": true, "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903213.546, "dur": 7.411,
    "args": {
      "Python parent id": 2762, "Python id": 2767, "finished": true, "Ev Idx": 5667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903229.886, "dur": 8.773,
    "args": {
      "Python parent id": 2762, "Python id": 2768, "finished": true, "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903259.518, "dur": 21.168,
    "args": {
      "Python parent id": 12, "Python id": 2769, "finished": true, "Ev Idx": 5669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903260.035, "dur": 14.218,
    "args": {
      "Python parent id": 2769, "Python id": 2770, "finished": true, "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903260.233, "dur": 0.477,
    "args": {
      "Python parent id": 2770, "Python id": 2771, "finished": true, "Ev Idx": 5671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903260.404, "dur": 0.214,
    "args": {
      "Python parent id": 2771, "Python id": 2772, "finished": true, "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903260.512, "dur": 0.070,
    "args": {
      "Python parent id": 2772, "Python id": 2773, "finished": true, "Ev Idx": 5673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903261.136, "dur": 10.435,
    "args": {
      "Python parent id": 2770, "Python id": 2774, "finished": true, "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903261.252, "dur": 0.118,
    "args": {
      "Python parent id": 2774, "Python id": 2775, "finished": true, "Ev Idx": 5675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903261.495, "dur": 0.063,
    "args": {
      "Python parent id": 2774, "Python id": 2776, "finished": true, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903261.709, "dur": 0.059,
    "args": {
      "Python parent id": 2774, "Python id": 2777, "finished": true, "Ev Idx": 5677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903262.315, "dur": 0.040,
    "args": {
      "Python parent id": 2774, "Python id": 2778, "finished": true, "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903262.687, "dur": 0.083,
    "args": {
      "Python parent id": 2774, "Python id": 2779, "finished": true, "Ev Idx": 5679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903263.148, "dur": 8.332,
    "args": {
      "Python parent id": 2774, "Python id": 2780, "finished": true, "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903263.258, "dur": 0.066,
    "args": {
      "Python parent id": 2780, "Python id": 2781, "finished": true, "Ev Idx": 5681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903263.478, "dur": 0.058,
    "args": {
      "Python parent id": 2780, "Python id": 2782, "finished": true, "Ev Idx": 5682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903263.771, "dur": 0.069,
    "args": {
      "Python parent id": 2780, "Python id": 2783, "finished": true, "Ev Idx": 5683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903264.263, "dur": 0.048,
    "args": {
      "Python parent id": 2780, "Python id": 2784, "finished": true, "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903264.707, "dur": 6.696,
    "args": {
      "Python parent id": 2780, "Python id": 2785, "finished": true, "Ev Idx": 5685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903264.978, "dur": 6.320,
    "args": {
      "Python parent id": 2785, "Python id": 2786, "finished": true, "Ev Idx": 5686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903265.224, "dur": 3.974,
    "args": {
      "Python parent id": 2786, "Python id": 2787, "finished": true, "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903265.447, "dur": 3.707,
    "args": {
      "Python parent id": 2787, "Python id": 2788, "finished": true, "Ev Idx": 5688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903265.622, "dur": 0.368,
    "args": {
      "Python parent id": 2788, "Python id": 2789, "finished": true, "Ev Idx": 5689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903265.805, "dur": 0.106,
    "args": {
      "Python parent id": 2789, "Python id": 2790, "finished": true, "Ev Idx": 5690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903266.171, "dur": 2.591,
    "args": {
      "Python parent id": 2788, "Python id": 2791, "finished": true, "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903266.439, "dur": 2.216,
    "args": {
      "Python parent id": 2791, "Python id": 2792, "finished": true, "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903266.698, "dur": 1.871,
    "args": {
      "Python parent id": 2792, "Python id": 2793, "finished": true, "Ev Idx": 5693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903266.874, "dur": 1.384,
    "args": {
      "Python parent id": 2793, "Python id": 2794, "finished": true, "Ev Idx": 5694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903267.088, "dur": 0.622,
    "args": {
      "Python parent id": 2794, "Python id": 2795, "finished": true, "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903267.212, "dur": 0.042,
    "args": {
      "Python parent id": 2795, "Python id": 2796, "finished": true, "Ev Idx": 5696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903267.429, "dur": 0.208,
    "args": {
      "Python parent id": 2795, "Python id": 2797, "finished": true, "Ev Idx": 5697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903268.957, "dur": 0.147,
    "args": {
      "Python parent id": 2788, "Python id": 2798, "finished": true, "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903269.412, "dur": 0.124,
    "args": {
      "Python parent id": 2786, "Python id": 2799, "finished": true, "Ev Idx": 5699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903269.903, "dur": 1.356,
    "args": {
      "Python parent id": 2786, "Python id": 2800, "finished": true, "Ev Idx": 5700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903270.078, "dur": 1.098,
    "args": {
      "Python parent id": 2800, "Python id": 2801, "finished": true, "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903270.231, "dur": 0.444,
    "args": {
      "Python parent id": 2801, "Python id": 2802, "finished": true, "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903270.405, "dur": 0.188,
    "args": {
      "Python parent id": 2802, "Python id": 2803, "finished": true, "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903270.503, "dur": 0.058,
    "args": {
      "Python parent id": 2803, "Python id": 2804, "finished": true, "Ev Idx": 5704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903270.828, "dur": 0.300,
    "args": {
      "Python parent id": 2801, "Python id": 2805, "finished": true, "Ev Idx": 5705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903271.722, "dur": 0.258,
    "args": {
      "Python parent id": 2770, "Python id": 2806, "finished": true, "Ev Idx": 5706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903272.729, "dur": 1.338,
    "args": {
      "Python parent id": 2770, "Python id": 2807, "finished": true, "Ev Idx": 5707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903273.409, "dur": 0.606,
    "args": {
      "Python parent id": 2807, "Python id": 2808, "finished": true, "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903274.555, "dur": 6.047,
    "args": {
      "Python parent id": 2769, "Python id": 2809, "finished": true, "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903281.318, "dur": 8.752,
    "args": {
      "Python parent id": 12, "Python id": 2810, "finished": true, "Ev Idx": 5710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903290.601, "dur": 0.116,
    "args": {
      "Python parent id": 12, "Python id": 2811, "finished": true, "Ev Idx": 5711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903291.378, "dur": 0.214,
    "args": {
      "Python parent id": 12, "Python id": 2812, "finished": true, "Ev Idx": 5712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903292.069, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 2813, "finished": true, "Ev Idx": 5713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903292.598, "dur": 10.779,
    "args": {
      "Python parent id": 12, "Python id": 2814, "finished": true, "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903304.959, "dur": 17.371,
    "args": {
      "Python parent id": 12, "Python id": 2815, "finished": true, "Ev Idx": 5715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903305.288, "dur": 12.936,
    "args": {
      "Python parent id": 2815, "Python id": 2816, "finished": true, "Ev Idx": 5716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903305.534, "dur": 0.770,
    "args": {
      "Python parent id": 2816, "Python id": 2817, "finished": true, "Ev Idx": 5717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903305.716, "dur": 0.511,
    "args": {
      "Python parent id": 2817, "Python id": 2818, "finished": true, "Ev Idx": 5718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903306.029, "dur": 0.142,
    "args": {
      "Python parent id": 2818, "Python id": 2819, "finished": true, "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903306.682, "dur": 9.390,
    "args": {
      "Python parent id": 2816, "Python id": 2820, "finished": true, "Ev Idx": 5720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903306.810, "dur": 0.098,
    "args": {
      "Python parent id": 2820, "Python id": 2821, "finished": true, "Ev Idx": 5721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903307.028, "dur": 0.082,
    "args": {
      "Python parent id": 2820, "Python id": 2822, "finished": true, "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903307.297, "dur": 0.054,
    "args": {
      "Python parent id": 2820, "Python id": 2823, "finished": true, "Ev Idx": 5723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903307.800, "dur": 0.058,
    "args": {
      "Python parent id": 2820, "Python id": 2824, "finished": true, "Ev Idx": 5724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903308.201, "dur": 0.069,
    "args": {
      "Python parent id": 2820, "Python id": 2825, "finished": true, "Ev Idx": 5725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903308.699, "dur": 7.305,
    "args": {
      "Python parent id": 2820, "Python id": 2826, "finished": true, "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903308.793, "dur": 0.064,
    "args": {
      "Python parent id": 2826, "Python id": 2827, "finished": true, "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903309.024, "dur": 0.059,
    "args": {
      "Python parent id": 2826, "Python id": 2828, "finished": true, "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903309.272, "dur": 0.065,
    "args": {
      "Python parent id": 2826, "Python id": 2829, "finished": true, "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903309.730, "dur": 0.078,
    "args": {
      "Python parent id": 2826, "Python id": 2830, "finished": true, "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903310.165, "dur": 5.782,
    "args": {
      "Python parent id": 2826, "Python id": 2831, "finished": true, "Ev Idx": 5731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903310.504, "dur": 5.359,
    "args": {
      "Python parent id": 2831, "Python id": 2832, "finished": true, "Ev Idx": 5732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903310.638, "dur": 3.290,
    "args": {
      "Python parent id": 2832, "Python id": 2833, "finished": true, "Ev Idx": 5733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903310.859, "dur": 3.024,
    "args": {
      "Python parent id": 2833, "Python id": 2834, "finished": true, "Ev Idx": 5734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903311.011, "dur": 0.308,
    "args": {
      "Python parent id": 2834, "Python id": 2835, "finished": true, "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903311.201, "dur": 0.066,
    "args": {
      "Python parent id": 2835, "Python id": 2836, "finished": true, "Ev Idx": 5736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903311.514, "dur": 2.041,
    "args": {
      "Python parent id": 2834, "Python id": 2837, "finished": true, "Ev Idx": 5737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903311.758, "dur": 1.699,
    "args": {
      "Python parent id": 2837, "Python id": 2838, "finished": true, "Ev Idx": 5738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903311.914, "dur": 1.475,
    "args": {
      "Python parent id": 2838, "Python id": 2839, "finished": true, "Ev Idx": 5739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903312.042, "dur": 1.111,
    "args": {
      "Python parent id": 2839, "Python id": 2840, "finished": true, "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903312.204, "dur": 0.485,
    "args": {
      "Python parent id": 2840, "Python id": 2841, "finished": true, "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903312.289, "dur": 0.050,
    "args": {
      "Python parent id": 2841, "Python id": 2842, "finished": true, "Ev Idx": 5742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903312.514, "dur": 0.130,
    "args": {
      "Python parent id": 2841, "Python id": 2843, "finished": true, "Ev Idx": 5743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903313.742, "dur": 0.086,
    "args": {
      "Python parent id": 2834, "Python id": 2844, "finished": true, "Ev Idx": 5744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903314.107, "dur": 0.110,
    "args": {
      "Python parent id": 2832, "Python id": 2845, "finished": true, "Ev Idx": 5745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903314.480, "dur": 1.311,
    "args": {
      "Python parent id": 2832, "Python id": 2846, "finished": true, "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903314.622, "dur": 1.124,
    "args": {
      "Python parent id": 2846, "Python id": 2847, "finished": true, "Ev Idx": 5747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903314.760, "dur": 0.557,
    "args": {
      "Python parent id": 2847, "Python id": 2848, "finished": true, "Ev Idx": 5748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903314.901, "dur": 0.352,
    "args": {
      "Python parent id": 2848, "Python id": 2849, "finished": true, "Ev Idx": 5749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903315.124, "dur": 0.074,
    "args": {
      "Python parent id": 2849, "Python id": 2850, "finished": true, "Ev Idx": 5750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903315.478, "dur": 0.238,
    "args": {
      "Python parent id": 2847, "Python id": 2851, "finished": true, "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903316.202, "dur": 0.203,
    "args": {
      "Python parent id": 2816, "Python id": 2852, "finished": true, "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903317.019, "dur": 1.049,
    "args": {
      "Python parent id": 2816, "Python id": 2853, "finished": true, "Ev Idx": 5753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903317.431, "dur": 0.588,
    "args": {
      "Python parent id": 2853, "Python id": 2854, "finished": true, "Ev Idx": 5754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903318.537, "dur": 3.748,
    "args": {
      "Python parent id": 2815, "Python id": 2855, "finished": true, "Ev Idx": 5755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903323.351, "dur": 100.762,
    "args": {
      "Python parent id": 12, "Python id": 2856, "finished": true, "Ev Idx": 5756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903323.770, "dur": 0.409,
    "args": {
      "Python parent id": 2856, "Python id": 2857, "finished": true, "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903324.064, "dur": 0.048,
    "args": {
      "Python parent id": 2857, "Python id": 2858, "finished": true, "Ev Idx": 5758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903324.568, "dur": 0.417,
    "args": {
      "Python parent id": 2856, "Python id": 2859, "finished": true, "Ev Idx": 5759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903325.227, "dur": 0.255,
    "args": {
      "Python parent id": 2856, "Python id": 2860, "finished": true, "Ev Idx": 5760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903325.752, "dur": 1.778,
    "args": {
      "Python parent id": 2856, "Python id": 2861, "finished": true, "Ev Idx": 5761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903328.200, "dur": 95.805,
    "args": {
      "Python parent id": 2856, "Python id": 2862, "finished": true, "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903329.191, "dur": 94.634,
    "args": {
      "Python parent id": 2862, "Python id": 2863, "finished": true, "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903329.479, "dur": 94.235,
    "args": {
      "Python parent id": 2863, "Python id": 2864, "finished": true, "Ev Idx": 5764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903335.084, "dur": 87.186,
    "args": {
      "Python parent id": 2864, "Python id": 2865, "finished": true, "Ev Idx": 5765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903335.435, "dur": 86.753,
    "args": {
      "Python parent id": 2865, "Python id": 2866, "finished": true, "Ev Idx": 5766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903336.064, "dur": 85.959,
    "args": {
      "Python parent id": 2866, "Python id": 2867, "finished": true, "Ev Idx": 5767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903336.715, "dur": 0.491,
    "args": {
      "Python parent id": 2867, "Python id": 2868, "finished": true, "Ev Idx": 5768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903337.066, "dur": 0.070,
    "args": {
      "Python parent id": 2868, "Python id": 2869, "finished": true, "Ev Idx": 5769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903337.578, "dur": 0.376,
    "args": {
      "Python parent id": 2867, "Python id": 2870, "finished": true, "Ev Idx": 5770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903337.827, "dur": 0.079,
    "args": {
      "Python parent id": 2870, "Python id": 2871, "finished": true, "Ev Idx": 5771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903338.335, "dur": 83.576,
    "args": {
      "Python parent id": 2867, "Python id": 2872, "finished": true, "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903339.159, "dur": 81.409,
    "args": {
      "Python parent id": 2872, "Python id": 2873, "finished": true, "Ev Idx": 5773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903348.495, "dur": 20.765,
    "args": {
      "Python parent id": 2873, "Python id": 2874, "finished": true, "Ev Idx": 5774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903348.617, "dur": 0.144,
    "args": {
      "Python parent id": 2874, "Python id": 2875, "finished": true, "Ev Idx": 5775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903349.107, "dur": 20.021,
    "args": {
      "Python parent id": 2874, "Python id": 2876, "finished": true, "Ev Idx": 5776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903349.286, "dur": 8.905,
    "args": {
      "Python parent id": 2876, "Python id": 2877, "finished": true, "Ev Idx": 5777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903378.629, "dur": 7.639,
    "args": {
      "Python parent id": 2873, "Python id": 2878, "finished": true, "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903395.619, "dur": 9.622,
    "args": {
      "Python parent id": 2873, "Python id": 2879, "finished": true, "Ev Idx": 5779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903425.476, "dur": 19.454,
    "args": {
      "Python parent id": 12, "Python id": 2880, "finished": true, "Ev Idx": 5780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903425.856, "dur": 14.213,
    "args": {
      "Python parent id": 2880, "Python id": 2881, "finished": true, "Ev Idx": 5781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903426.075, "dur": 0.790,
    "args": {
      "Python parent id": 2881, "Python id": 2882, "finished": true, "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903426.341, "dur": 0.453,
    "args": {
      "Python parent id": 2882, "Python id": 2883, "finished": true, "Ev Idx": 5783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903426.624, "dur": 0.111,
    "args": {
      "Python parent id": 2883, "Python id": 2884, "finished": true, "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903427.254, "dur": 10.269,
    "args": {
      "Python parent id": 2881, "Python id": 2885, "finished": true, "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903427.348, "dur": 0.068,
    "args": {
      "Python parent id": 2885, "Python id": 2886, "finished": true, "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903427.589, "dur": 0.063,
    "args": {
      "Python parent id": 2885, "Python id": 2887, "finished": true, "Ev Idx": 5787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903427.894, "dur": 0.047,
    "args": {
      "Python parent id": 2885, "Python id": 2888, "finished": true, "Ev Idx": 5788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903428.373, "dur": 0.073,
    "args": {
      "Python parent id": 2885, "Python id": 2889, "finished": true, "Ev Idx": 5789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903428.749, "dur": 0.089,
    "args": {
      "Python parent id": 2885, "Python id": 2890, "finished": true, "Ev Idx": 5790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903429.261, "dur": 8.196,
    "args": {
      "Python parent id": 2885, "Python id": 2891, "finished": true, "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903429.386, "dur": 0.065,
    "args": {
      "Python parent id": 2891, "Python id": 2892, "finished": true, "Ev Idx": 5792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903429.622, "dur": 0.065,
    "args": {
      "Python parent id": 2891, "Python id": 2893, "finished": true, "Ev Idx": 5793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903429.889, "dur": 0.080,
    "args": {
      "Python parent id": 2891, "Python id": 2894, "finished": true, "Ev Idx": 5794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903430.377, "dur": 0.040,
    "args": {
      "Python parent id": 2891, "Python id": 2895, "finished": true, "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903430.799, "dur": 6.562,
    "args": {
      "Python parent id": 2891, "Python id": 2896, "finished": true, "Ev Idx": 5796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903431.150, "dur": 6.155,
    "args": {
      "Python parent id": 2896, "Python id": 2897, "finished": true, "Ev Idx": 5797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903431.418, "dur": 3.819,
    "args": {
      "Python parent id": 2897, "Python id": 2898, "finished": true, "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903431.650, "dur": 3.523,
    "args": {
      "Python parent id": 2898, "Python id": 2899, "finished": true, "Ev Idx": 5799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903431.829, "dur": 0.336,
    "args": {
      "Python parent id": 2899, "Python id": 2900, "finished": true, "Ev Idx": 5800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903432.001, "dur": 0.067,
    "args": {
      "Python parent id": 2900, "Python id": 2901, "finished": true, "Ev Idx": 5801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903432.328, "dur": 2.425,
    "args": {
      "Python parent id": 2899, "Python id": 2902, "finished": true, "Ev Idx": 5802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903432.609, "dur": 2.040,
    "args": {
      "Python parent id": 2902, "Python id": 2903, "finished": true, "Ev Idx": 5803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903432.828, "dur": 1.738,
    "args": {
      "Python parent id": 2903, "Python id": 2904, "finished": true, "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903432.989, "dur": 1.354,
    "args": {
      "Python parent id": 2904, "Python id": 2905, "finished": true, "Ev Idx": 5805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903433.220, "dur": 0.577,
    "args": {
      "Python parent id": 2905, "Python id": 2906, "finished": true, "Ev Idx": 5806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903433.341, "dur": 0.039,
    "args": {
      "Python parent id": 2906, "Python id": 2907, "finished": true, "Ev Idx": 5807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903433.594, "dur": 0.154,
    "args": {
      "Python parent id": 2906, "Python id": 2908, "finished": true, "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903434.960, "dur": 0.152,
    "args": {
      "Python parent id": 2899, "Python id": 2909, "finished": true, "Ev Idx": 5809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903435.421, "dur": 0.093,
    "args": {
      "Python parent id": 2897, "Python id": 2910, "finished": true, "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903435.838, "dur": 1.410,
    "args": {
      "Python parent id": 2897, "Python id": 2911, "finished": true, "Ev Idx": 5811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903435.984, "dur": 1.220,
    "args": {
      "Python parent id": 2911, "Python id": 2912, "finished": true, "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903436.124, "dur": 0.463,
    "args": {
      "Python parent id": 2912, "Python id": 2913, "finished": true, "Ev Idx": 5813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903436.267, "dur": 0.263,
    "args": {
      "Python parent id": 2913, "Python id": 2914, "finished": true, "Ev Idx": 5814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903436.429, "dur": 0.049,
    "args": {
      "Python parent id": 2914, "Python id": 2915, "finished": true, "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903436.813, "dur": 0.339,
    "args": {
      "Python parent id": 2912, "Python id": 2916, "finished": true, "Ev Idx": 5816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903437.647, "dur": 0.270,
    "args": {
      "Python parent id": 2881, "Python id": 2917, "finished": true, "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903438.573, "dur": 1.306,
    "args": {
      "Python parent id": 2881, "Python id": 2918, "finished": true, "Ev Idx": 5818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903439.153, "dur": 0.656,
    "args": {
      "Python parent id": 2918, "Python id": 2919, "finished": true, "Ev Idx": 5819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903440.412, "dur": 4.434,
    "args": {
      "Python parent id": 2880, "Python id": 2920, "finished": true, "Ev Idx": 5820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903445.597, "dur": 8.253,
    "args": {
      "Python parent id": 12, "Python id": 2921, "finished": true, "Ev Idx": 5821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903454.277, "dur": 0.142,
    "args": {
      "Python parent id": 12, "Python id": 2922, "finished": true, "Ev Idx": 5822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903455.086, "dur": 0.188,
    "args": {
      "Python parent id": 12, "Python id": 2923, "finished": true, "Ev Idx": 5823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903455.781, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 2924, "finished": true, "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903456.258, "dur": 10.591,
    "args": {
      "Python parent id": 12, "Python id": 2925, "finished": true, "Ev Idx": 5825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903468.236, "dur": 15.733,
    "args": {
      "Python parent id": 12, "Python id": 2926, "finished": true, "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903468.598, "dur": 11.800,
    "args": {
      "Python parent id": 2926, "Python id": 2927, "finished": true, "Ev Idx": 5827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903468.754, "dur": 0.584,
    "args": {
      "Python parent id": 2927, "Python id": 2928, "finished": true, "Ev Idx": 5828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903468.914, "dur": 0.348,
    "args": {
      "Python parent id": 2928, "Python id": 2929, "finished": true, "Ev Idx": 5829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903469.069, "dur": 0.156,
    "args": {
      "Python parent id": 2929, "Python id": 2930, "finished": true, "Ev Idx": 5830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903469.630, "dur": 8.824,
    "args": {
      "Python parent id": 2927, "Python id": 2931, "finished": true, "Ev Idx": 5831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903469.724, "dur": 0.069,
    "args": {
      "Python parent id": 2931, "Python id": 2932, "finished": true, "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903469.911, "dur": 0.064,
    "args": {
      "Python parent id": 2931, "Python id": 2933, "finished": true, "Ev Idx": 5833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903470.127, "dur": 0.064,
    "args": {
      "Python parent id": 2931, "Python id": 2934, "finished": true, "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903470.623, "dur": 0.046,
    "args": {
      "Python parent id": 2931, "Python id": 2935, "finished": true, "Ev Idx": 5835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903470.990, "dur": 0.073,
    "args": {
      "Python parent id": 2931, "Python id": 2936, "finished": true, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903471.427, "dur": 6.957,
    "args": {
      "Python parent id": 2931, "Python id": 2937, "finished": true, "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903471.514, "dur": 0.061,
    "args": {
      "Python parent id": 2937, "Python id": 2938, "finished": true, "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903471.742, "dur": 0.061,
    "args": {
      "Python parent id": 2937, "Python id": 2939, "finished": true, "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903471.975, "dur": 0.065,
    "args": {
      "Python parent id": 2937, "Python id": 2940, "finished": true, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903472.396, "dur": 0.038,
    "args": {
      "Python parent id": 2937, "Python id": 2941, "finished": true, "Ev Idx": 5841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903472.793, "dur": 5.478,
    "args": {
      "Python parent id": 2937, "Python id": 2942, "finished": true, "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903473.001, "dur": 5.209,
    "args": {
      "Python parent id": 2942, "Python id": 2943, "finished": true, "Ev Idx": 5843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903473.147, "dur": 3.291,
    "args": {
      "Python parent id": 2943, "Python id": 2944, "finished": true, "Ev Idx": 5844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903473.371, "dur": 3.019,
    "args": {
      "Python parent id": 2944, "Python id": 2945, "finished": true, "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903473.524, "dur": 0.292,
    "args": {
      "Python parent id": 2945, "Python id": 2946, "finished": true, "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903473.682, "dur": 0.065,
    "args": {
      "Python parent id": 2946, "Python id": 2947, "finished": true, "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.004, "dur": 2.046,
    "args": {
      "Python parent id": 2945, "Python id": 2948, "finished": true, "Ev Idx": 5848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.219, "dur": 1.754,
    "args": {
      "Python parent id": 2948, "Python id": 2949, "finished": true, "Ev Idx": 5849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.426, "dur": 1.477,
    "args": {
      "Python parent id": 2949, "Python id": 2950, "finished": true, "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.563, "dur": 1.111,
    "args": {
      "Python parent id": 2950, "Python id": 2951, "finished": true, "Ev Idx": 5851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.739, "dur": 0.484,
    "args": {
      "Python parent id": 2951, "Python id": 2952, "finished": true, "Ev Idx": 5852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903474.831, "dur": 0.053,
    "args": {
      "Python parent id": 2952, "Python id": 2953, "finished": true, "Ev Idx": 5853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903475.043, "dur": 0.134,
    "args": {
      "Python parent id": 2952, "Python id": 2954, "finished": true, "Ev Idx": 5854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903476.263, "dur": 0.071,
    "args": {
      "Python parent id": 2945, "Python id": 2955, "finished": true, "Ev Idx": 5855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903476.621, "dur": 0.084,
    "args": {
      "Python parent id": 2943, "Python id": 2956, "finished": true, "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903476.972, "dur": 1.193,
    "args": {
      "Python parent id": 2943, "Python id": 2957, "finished": true, "Ev Idx": 5857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903477.131, "dur": 0.991,
    "args": {
      "Python parent id": 2957, "Python id": 2958, "finished": true, "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903477.263, "dur": 0.410,
    "args": {
      "Python parent id": 2958, "Python id": 2959, "finished": true, "Ev Idx": 5859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903477.411, "dur": 0.201,
    "args": {
      "Python parent id": 2959, "Python id": 2960, "finished": true, "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903477.523, "dur": 0.058,
    "args": {
      "Python parent id": 2960, "Python id": 2961, "finished": true, "Ev Idx": 5861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903477.838, "dur": 0.233,
    "args": {
      "Python parent id": 2958, "Python id": 2962, "finished": true, "Ev Idx": 5862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903478.598, "dur": 0.156,
    "args": {
      "Python parent id": 2927, "Python id": 2963, "finished": true, "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903479.363, "dur": 0.911,
    "args": {
      "Python parent id": 2927, "Python id": 2964, "finished": true, "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903479.714, "dur": 0.508,
    "args": {
      "Python parent id": 2964, "Python id": 2965, "finished": true, "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903480.662, "dur": 3.263,
    "args": {
      "Python parent id": 2926, "Python id": 2966, "finished": true, "Ev Idx": 5866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903484.845, "dur": 98.771,
    "args": {
      "Python parent id": 12, "Python id": 2967, "finished": true, "Ev Idx": 5867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903485.304, "dur": 0.408,
    "args": {
      "Python parent id": 2967, "Python id": 2968, "finished": true, "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903485.571, "dur": 0.074,
    "args": {
      "Python parent id": 2968, "Python id": 2969, "finished": true, "Ev Idx": 5869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903486.030, "dur": 0.342,
    "args": {
      "Python parent id": 2967, "Python id": 2970, "finished": true, "Ev Idx": 5870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903486.589, "dur": 0.297,
    "args": {
      "Python parent id": 2967, "Python id": 2971, "finished": true, "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903487.128, "dur": 1.760,
    "args": {
      "Python parent id": 2967, "Python id": 2972, "finished": true, "Ev Idx": 5872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903489.567, "dur": 93.921,
    "args": {
      "Python parent id": 2967, "Python id": 2973, "finished": true, "Ev Idx": 5873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903490.443, "dur": 92.814,
    "args": {
      "Python parent id": 2973, "Python id": 2974, "finished": true, "Ev Idx": 5874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903490.731, "dur": 92.391,
    "args": {
      "Python parent id": 2974, "Python id": 2975, "finished": true, "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903498.253, "dur": 83.473,
    "args": {
      "Python parent id": 2975, "Python id": 2976, "finished": true, "Ev Idx": 5876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903498.569, "dur": 83.085,
    "args": {
      "Python parent id": 2976, "Python id": 2977, "finished": true, "Ev Idx": 5877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903499.166, "dur": 82.267,
    "args": {
      "Python parent id": 2977, "Python id": 2978, "finished": true, "Ev Idx": 5878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903499.767, "dur": 0.477,
    "args": {
      "Python parent id": 2978, "Python id": 2979, "finished": true, "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903500.079, "dur": 0.084,
    "args": {
      "Python parent id": 2979, "Python id": 2980, "finished": true, "Ev Idx": 5880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903500.582, "dur": 0.346,
    "args": {
      "Python parent id": 2978, "Python id": 2981, "finished": true, "Ev Idx": 5881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903500.814, "dur": 0.058,
    "args": {
      "Python parent id": 2981, "Python id": 2982, "finished": true, "Ev Idx": 5882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903501.279, "dur": 80.052,
    "args": {
      "Python parent id": 2978, "Python id": 2983, "finished": true, "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903501.909, "dur": 78.153,
    "args": {
      "Python parent id": 2983, "Python id": 2984, "finished": true, "Ev Idx": 5884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903511.395, "dur": 17.524,
    "args": {
      "Python parent id": 2984, "Python id": 2985, "finished": true, "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903511.529, "dur": 0.140,
    "args": {
      "Python parent id": 2985, "Python id": 2986, "finished": true, "Ev Idx": 5886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903511.953, "dur": 16.857,
    "args": {
      "Python parent id": 2985, "Python id": 2987, "finished": true, "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903512.125, "dur": 7.828,
    "args": {
      "Python parent id": 2987, "Python id": 2988, "finished": true, "Ev Idx": 5888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903537.107, "dur": 9.006,
    "args": {
      "Python parent id": 2984, "Python id": 2989, "finished": true, "Ev Idx": 5889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903556.238, "dur": 8.550,
    "args": {
      "Python parent id": 2984, "Python id": 2990, "finished": true, "Ev Idx": 5890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903585.450, "dur": 18.718,
    "args": {
      "Python parent id": 12, "Python id": 2991, "finished": true, "Ev Idx": 5891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903585.884, "dur": 13.394,
    "args": {
      "Python parent id": 2991, "Python id": 2992, "finished": true, "Ev Idx": 5892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903586.072, "dur": 0.639,
    "args": {
      "Python parent id": 2992, "Python id": 2993, "finished": true, "Ev Idx": 5893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903586.311, "dur": 0.334,
    "args": {
      "Python parent id": 2993, "Python id": 2994, "finished": true, "Ev Idx": 5894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903586.471, "dur": 0.109,
    "args": {
      "Python parent id": 2994, "Python id": 2995, "finished": true, "Ev Idx": 5895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903587.052, "dur": 9.776,
    "args": {
      "Python parent id": 2992, "Python id": 2996, "finished": true, "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903587.158, "dur": 0.066,
    "args": {
      "Python parent id": 2996, "Python id": 2997, "finished": true, "Ev Idx": 5897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903587.354, "dur": 0.062,
    "args": {
      "Python parent id": 2996, "Python id": 2998, "finished": true, "Ev Idx": 5898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903587.578, "dur": 0.059,
    "args": {
      "Python parent id": 2996, "Python id": 2999, "finished": true, "Ev Idx": 5899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903588.167, "dur": 0.053,
    "args": {
      "Python parent id": 2996, "Python id": 3000, "finished": true, "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903588.492, "dur": 0.072,
    "args": {
      "Python parent id": 2996, "Python id": 3001, "finished": true, "Ev Idx": 5901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903588.943, "dur": 7.817,
    "args": {
      "Python parent id": 2996, "Python id": 3002, "finished": true, "Ev Idx": 5902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903589.049, "dur": 0.069,
    "args": {
      "Python parent id": 3002, "Python id": 3003, "finished": true, "Ev Idx": 5903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903589.309, "dur": 0.058,
    "args": {
      "Python parent id": 3002, "Python id": 3004, "finished": true, "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903589.573, "dur": 0.068,
    "args": {
      "Python parent id": 3002, "Python id": 3005, "finished": true, "Ev Idx": 5905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903590.043, "dur": 0.052,
    "args": {
      "Python parent id": 3002, "Python id": 3006, "finished": true, "Ev Idx": 5906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903590.405, "dur": 6.266,
    "args": {
      "Python parent id": 3002, "Python id": 3007, "finished": true, "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903590.647, "dur": 5.939,
    "args": {
      "Python parent id": 3007, "Python id": 3008, "finished": true, "Ev Idx": 5908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903590.769, "dur": 3.706,
    "args": {
      "Python parent id": 3008, "Python id": 3009, "finished": true, "Ev Idx": 5909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903590.951, "dur": 3.458,
    "args": {
      "Python parent id": 3009, "Python id": 3010, "finished": true, "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903591.147, "dur": 0.382,
    "args": {
      "Python parent id": 3010, "Python id": 3011, "finished": true, "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903591.367, "dur": 0.080,
    "args": {
      "Python parent id": 3011, "Python id": 3012, "finished": true, "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903591.720, "dur": 2.296,
    "args": {
      "Python parent id": 3010, "Python id": 3013, "finished": true, "Ev Idx": 5913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903591.967, "dur": 1.936,
    "args": {
      "Python parent id": 3013, "Python id": 3014, "finished": true, "Ev Idx": 5914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903592.198, "dur": 1.618,
    "args": {
      "Python parent id": 3014, "Python id": 3015, "finished": true, "Ev Idx": 5915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903592.342, "dur": 1.275,
    "args": {
      "Python parent id": 3015, "Python id": 3016, "finished": true, "Ev Idx": 5916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903592.542, "dur": 0.582,
    "args": {
      "Python parent id": 3016, "Python id": 3017, "finished": true, "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903592.640, "dur": 0.046,
    "args": {
      "Python parent id": 3017, "Python id": 3018, "finished": true, "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903592.882, "dur": 0.181,
    "args": {
      "Python parent id": 3017, "Python id": 3019, "finished": true, "Ev Idx": 5919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903594.219, "dur": 0.144,
    "args": {
      "Python parent id": 3010, "Python id": 3020, "finished": true, "Ev Idx": 5920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903594.715, "dur": 0.125,
    "args": {
      "Python parent id": 3008, "Python id": 3021, "finished": true, "Ev Idx": 5921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903595.184, "dur": 1.355,
    "args": {
      "Python parent id": 3008, "Python id": 3022, "finished": true, "Ev Idx": 5922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903595.382, "dur": 1.115,
    "args": {
      "Python parent id": 3022, "Python id": 3023, "finished": true, "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903595.531, "dur": 0.440,
    "args": {
      "Python parent id": 3023, "Python id": 3024, "finished": true, "Ev Idx": 5924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903595.679, "dur": 0.209,
    "args": {
      "Python parent id": 3024, "Python id": 3025, "finished": true, "Ev Idx": 5925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903595.791, "dur": 0.063,
    "args": {
      "Python parent id": 3025, "Python id": 3026, "finished": true, "Ev Idx": 5926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903596.118, "dur": 0.310,
    "args": {
      "Python parent id": 3023, "Python id": 3027, "finished": true, "Ev Idx": 5927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903596.980, "dur": 0.280,
    "args": {
      "Python parent id": 2992, "Python id": 3028, "finished": true, "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903597.954, "dur": 1.182,
    "args": {
      "Python parent id": 2992, "Python id": 3029, "finished": true, "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903598.509, "dur": 0.574,
    "args": {
      "Python parent id": 3029, "Python id": 3030, "finished": true, "Ev Idx": 5930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903599.634, "dur": 4.459,
    "args": {
      "Python parent id": 2991, "Python id": 3031, "finished": true, "Ev Idx": 5931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903604.783, "dur": 7.458,
    "args": {
      "Python parent id": 12, "Python id": 3032, "finished": true, "Ev Idx": 5932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903612.721, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 3033, "finished": true, "Ev Idx": 5933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903613.426, "dur": 0.203,
    "args": {
      "Python parent id": 12, "Python id": 3034, "finished": true, "Ev Idx": 5934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903614.191, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 3035, "finished": true, "Ev Idx": 5935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903614.693, "dur": 10.343,
    "args": {
      "Python parent id": 12, "Python id": 3036, "finished": true, "Ev Idx": 5936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903626.381, "dur": 15.811,
    "args": {
      "Python parent id": 12, "Python id": 3037, "finished": true, "Ev Idx": 5937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903626.741, "dur": 11.988,
    "args": {
      "Python parent id": 3037, "Python id": 3038, "finished": true, "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903626.898, "dur": 0.524,
    "args": {
      "Python parent id": 3038, "Python id": 3039, "finished": true, "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903627.093, "dur": 0.231,
    "args": {
      "Python parent id": 3039, "Python id": 3040, "finished": true, "Ev Idx": 5940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903627.177, "dur": 0.111,
    "args": {
      "Python parent id": 3040, "Python id": 3041, "finished": true, "Ev Idx": 5941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903627.733, "dur": 9.033,
    "args": {
      "Python parent id": 3038, "Python id": 3042, "finished": true, "Ev Idx": 5942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903627.859, "dur": 0.069,
    "args": {
      "Python parent id": 3042, "Python id": 3043, "finished": true, "Ev Idx": 5943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903628.059, "dur": 0.065,
    "args": {
      "Python parent id": 3042, "Python id": 3044, "finished": true, "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903628.308, "dur": 0.061,
    "args": {
      "Python parent id": 3042, "Python id": 3045, "finished": true, "Ev Idx": 5945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903628.871, "dur": 0.042,
    "args": {
      "Python parent id": 3042, "Python id": 3046, "finished": true, "Ev Idx": 5946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903629.196, "dur": 0.078,
    "args": {
      "Python parent id": 3042, "Python id": 3047, "finished": true, "Ev Idx": 5947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903629.658, "dur": 7.039,
    "args": {
      "Python parent id": 3042, "Python id": 3048, "finished": true, "Ev Idx": 5948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903629.759, "dur": 0.064,
    "args": {
      "Python parent id": 3048, "Python id": 3049, "finished": true, "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903629.973, "dur": 0.059,
    "args": {
      "Python parent id": 3048, "Python id": 3050, "finished": true, "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903630.188, "dur": 0.075,
    "args": {
      "Python parent id": 3048, "Python id": 3051, "finished": true, "Ev Idx": 5951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903630.614, "dur": 0.067,
    "args": {
      "Python parent id": 3048, "Python id": 3052, "finished": true, "Ev Idx": 5952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.019, "dur": 5.591,
    "args": {
      "Python parent id": 3048, "Python id": 3053, "finished": true, "Ev Idx": 5953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.210, "dur": 5.333,
    "args": {
      "Python parent id": 3053, "Python id": 3054, "finished": true, "Ev Idx": 5954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.352, "dur": 3.437,
    "args": {
      "Python parent id": 3054, "Python id": 3055, "finished": true, "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.576, "dur": 3.167,
    "args": {
      "Python parent id": 3055, "Python id": 3056, "finished": true, "Ev Idx": 5956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.715, "dur": 0.283,
    "args": {
      "Python parent id": 3056, "Python id": 3057, "finished": true, "Ev Idx": 5957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903631.870, "dur": 0.071,
    "args": {
      "Python parent id": 3057, "Python id": 3058, "finished": true, "Ev Idx": 5958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903632.151, "dur": 2.308,
    "args": {
      "Python parent id": 3056, "Python id": 3059, "finished": true, "Ev Idx": 5959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903632.392, "dur": 1.978,
    "args": {
      "Python parent id": 3059, "Python id": 3060, "finished": true, "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903632.620, "dur": 1.653,
    "args": {
      "Python parent id": 3060, "Python id": 3061, "finished": true, "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903632.789, "dur": 1.168,
    "args": {
      "Python parent id": 3061, "Python id": 3062, "finished": true, "Ev Idx": 5962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903633.001, "dur": 0.501,
    "args": {
      "Python parent id": 3062, "Python id": 3063, "finished": true, "Ev Idx": 5963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903633.088, "dur": 0.053,
    "args": {
      "Python parent id": 3063, "Python id": 3064, "finished": true, "Ev Idx": 5964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903633.324, "dur": 0.123,
    "args": {
      "Python parent id": 3063, "Python id": 3065, "finished": true, "Ev Idx": 5965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903634.641, "dur": 0.066,
    "args": {
      "Python parent id": 3056, "Python id": 3066, "finished": true, "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903634.955, "dur": 0.081,
    "args": {
      "Python parent id": 3054, "Python id": 3067, "finished": true, "Ev Idx": 5967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903635.293, "dur": 1.201,
    "args": {
      "Python parent id": 3054, "Python id": 3068, "finished": true, "Ev Idx": 5968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903635.436, "dur": 1.012,
    "args": {
      "Python parent id": 3068, "Python id": 3069, "finished": true, "Ev Idx": 5969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903635.580, "dur": 0.384,
    "args": {
      "Python parent id": 3069, "Python id": 3070, "finished": true, "Ev Idx": 5970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903635.724, "dur": 0.180,
    "args": {
      "Python parent id": 3070, "Python id": 3071, "finished": true, "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903635.814, "dur": 0.059,
    "args": {
      "Python parent id": 3071, "Python id": 3072, "finished": true, "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903636.096, "dur": 0.320,
    "args": {
      "Python parent id": 3069, "Python id": 3073, "finished": true, "Ev Idx": 5973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903636.889, "dur": 0.171,
    "args": {
      "Python parent id": 3038, "Python id": 3074, "finished": true, "Ev Idx": 5974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903637.681, "dur": 0.922,
    "args": {
      "Python parent id": 3038, "Python id": 3075, "finished": true, "Ev Idx": 5975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903638.076, "dur": 0.473,
    "args": {
      "Python parent id": 3075, "Python id": 3076, "finished": true, "Ev Idx": 5976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903638.996, "dur": 3.151,
    "args": {
      "Python parent id": 3037, "Python id": 3077, "finished": true, "Ev Idx": 5977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903643.130, "dur": 92.038,
    "args": {
      "Python parent id": 12, "Python id": 3078, "finished": true, "Ev Idx": 5978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903643.452, "dur": 0.443,
    "args": {
      "Python parent id": 3078, "Python id": 3079, "finished": true, "Ev Idx": 5979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903643.767, "dur": 0.076,
    "args": {
      "Python parent id": 3079, "Python id": 3080, "finished": true, "Ev Idx": 5980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903644.207, "dur": 0.407,
    "args": {
      "Python parent id": 3078, "Python id": 3081, "finished": true, "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903644.914, "dur": 0.267,
    "args": {
      "Python parent id": 3078, "Python id": 3082, "finished": true, "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903645.409, "dur": 1.664,
    "args": {
      "Python parent id": 3078, "Python id": 3083, "finished": true, "Ev Idx": 5983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903647.633, "dur": 87.416,
    "args": {
      "Python parent id": 3078, "Python id": 3084, "finished": true, "Ev Idx": 5984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903648.609, "dur": 86.193,
    "args": {
      "Python parent id": 3084, "Python id": 3085, "finished": true, "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903648.938, "dur": 85.726,
    "args": {
      "Python parent id": 3085, "Python id": 3086, "finished": true, "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903654.440, "dur": 78.750,
    "args": {
      "Python parent id": 3086, "Python id": 3087, "finished": true, "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903654.804, "dur": 78.272,
    "args": {
      "Python parent id": 3087, "Python id": 3088, "finished": true, "Ev Idx": 5988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903655.421, "dur": 77.522,
    "args": {
      "Python parent id": 3088, "Python id": 3089, "finished": true, "Ev Idx": 5989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903656.010, "dur": 1.269,
    "args": {
      "Python parent id": 3089, "Python id": 3090, "finished": true, "Ev Idx": 5990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903657.151, "dur": 0.047,
    "args": {
      "Python parent id": 3090, "Python id": 3091, "finished": true, "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903657.606, "dur": 0.357,
    "args": {
      "Python parent id": 3089, "Python id": 3092, "finished": true, "Ev Idx": 5992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903657.861, "dur": 0.047,
    "args": {
      "Python parent id": 3092, "Python id": 3093, "finished": true, "Ev Idx": 5993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903658.382, "dur": 74.444,
    "args": {
      "Python parent id": 3089, "Python id": 3094, "finished": true, "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903659.080, "dur": 72.444,
    "args": {
      "Python parent id": 3094, "Python id": 3095, "finished": true, "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903667.071, "dur": 18.093,
    "args": {
      "Python parent id": 3095, "Python id": 3096, "finished": true, "Ev Idx": 5996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903667.209, "dur": 0.085,
    "args": {
      "Python parent id": 3096, "Python id": 3097, "finished": true, "Ev Idx": 5997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903667.599, "dur": 17.471,
    "args": {
      "Python parent id": 3096, "Python id": 3098, "finished": true, "Ev Idx": 5998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903667.712, "dur": 7.689,
    "args": {
      "Python parent id": 3098, "Python id": 3099, "finished": true, "Ev Idx": 5999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903692.954, "dur": 6.622,
    "args": {
      "Python parent id": 3095, "Python id": 3100, "finished": true, "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903708.176, "dur": 7.844,
    "args": {
      "Python parent id": 3095, "Python id": 3101, "finished": true, "Ev Idx": 6001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903736.390, "dur": 18.343,
    "args": {
      "Python parent id": 12, "Python id": 3102, "finished": true, "Ev Idx": 6002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903736.778, "dur": 13.766,
    "args": {
      "Python parent id": 3102, "Python id": 3103, "finished": true, "Ev Idx": 6003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903736.986, "dur": 0.555,
    "args": {
      "Python parent id": 3103, "Python id": 3104, "finished": true, "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903737.186, "dur": 0.294,
    "args": {
      "Python parent id": 3104, "Python id": 3105, "finished": true, "Ev Idx": 6005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903737.340, "dur": 0.103,
    "args": {
      "Python parent id": 3105, "Python id": 3106, "finished": true, "Ev Idx": 6006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903737.847, "dur": 10.039,
    "args": {
      "Python parent id": 3103, "Python id": 3107, "finished": true, "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903737.982, "dur": 0.079,
    "args": {
      "Python parent id": 3107, "Python id": 3108, "finished": true, "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903738.189, "dur": 0.064,
    "args": {
      "Python parent id": 3107, "Python id": 3109, "finished": true, "Ev Idx": 6009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903738.459, "dur": 0.057,
    "args": {
      "Python parent id": 3107, "Python id": 3110, "finished": true, "Ev Idx": 6010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903738.977, "dur": 0.033,
    "args": {
      "Python parent id": 3107, "Python id": 3111, "finished": true, "Ev Idx": 6011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903739.306, "dur": 0.091,
    "args": {
      "Python parent id": 3107, "Python id": 3112, "finished": true, "Ev Idx": 6012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903739.746, "dur": 8.072,
    "args": {
      "Python parent id": 3107, "Python id": 3113, "finished": true, "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903739.836, "dur": 0.066,
    "args": {
      "Python parent id": 3113, "Python id": 3114, "finished": true, "Ev Idx": 6014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903740.090, "dur": 0.052,
    "args": {
      "Python parent id": 3113, "Python id": 3115, "finished": true, "Ev Idx": 6015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903740.302, "dur": 0.065,
    "args": {
      "Python parent id": 3113, "Python id": 3116, "finished": true, "Ev Idx": 6016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903740.757, "dur": 0.036,
    "args": {
      "Python parent id": 3113, "Python id": 3117, "finished": true, "Ev Idx": 6017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903741.188, "dur": 6.530,
    "args": {
      "Python parent id": 3113, "Python id": 3118, "finished": true, "Ev Idx": 6018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903741.471, "dur": 6.164,
    "args": {
      "Python parent id": 3118, "Python id": 3119, "finished": true, "Ev Idx": 6019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903741.702, "dur": 4.060,
    "args": {
      "Python parent id": 3119, "Python id": 3120, "finished": true, "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903741.897, "dur": 3.788,
    "args": {
      "Python parent id": 3120, "Python id": 3121, "finished": true, "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903742.089, "dur": 0.382,
    "args": {
      "Python parent id": 3121, "Python id": 3122, "finished": true, "Ev Idx": 6022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903742.309, "dur": 0.102,
    "args": {
      "Python parent id": 3122, "Python id": 3123, "finished": true, "Ev Idx": 6023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903742.706, "dur": 2.571,
    "args": {
      "Python parent id": 3121, "Python id": 3124, "finished": true, "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903742.998, "dur": 2.178,
    "args": {
      "Python parent id": 3124, "Python id": 3125, "finished": true, "Ev Idx": 6025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903743.216, "dur": 1.876,
    "args": {
      "Python parent id": 3125, "Python id": 3126, "finished": true, "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903743.410, "dur": 1.431,
    "args": {
      "Python parent id": 3126, "Python id": 3127, "finished": true, "Ev Idx": 6027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903743.614, "dur": 0.538,
    "args": {
      "Python parent id": 3127, "Python id": 3128, "finished": true, "Ev Idx": 6028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903743.705, "dur": 0.051,
    "args": {
      "Python parent id": 3128, "Python id": 3129, "finished": true, "Ev Idx": 6029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903743.952, "dur": 0.148,
    "args": {
      "Python parent id": 3128, "Python id": 3130, "finished": true, "Ev Idx": 6030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903745.494, "dur": 0.154,
    "args": {
      "Python parent id": 3121, "Python id": 3131, "finished": true, "Ev Idx": 6031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903745.925, "dur": 0.103,
    "args": {
      "Python parent id": 3119, "Python id": 3132, "finished": true, "Ev Idx": 6032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903746.365, "dur": 1.225,
    "args": {
      "Python parent id": 3119, "Python id": 3133, "finished": true, "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903746.509, "dur": 1.035,
    "args": {
      "Python parent id": 3133, "Python id": 3134, "finished": true, "Ev Idx": 6034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903746.670, "dur": 0.396,
    "args": {
      "Python parent id": 3134, "Python id": 3135, "finished": true, "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903746.809, "dur": 0.200,
    "args": {
      "Python parent id": 3135, "Python id": 3136, "finished": true, "Ev Idx": 6036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903746.920, "dur": 0.053,
    "args": {
      "Python parent id": 3136, "Python id": 3137, "finished": true, "Ev Idx": 6037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903747.190, "dur": 0.319,
    "args": {
      "Python parent id": 3134, "Python id": 3138, "finished": true, "Ev Idx": 6038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903748.037, "dur": 0.297,
    "args": {
      "Python parent id": 3103, "Python id": 3139, "finished": true, "Ev Idx": 6039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903749.092, "dur": 1.265,
    "args": {
      "Python parent id": 3103, "Python id": 3140, "finished": true, "Ev Idx": 6040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903749.634, "dur": 0.632,
    "args": {
      "Python parent id": 3140, "Python id": 3141, "finished": true, "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903750.871, "dur": 3.758,
    "args": {
      "Python parent id": 3102, "Python id": 3142, "finished": true, "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903755.441, "dur": 7.962,
    "args": {
      "Python parent id": 12, "Python id": 3143, "finished": true, "Ev Idx": 6043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903763.897, "dur": 0.137,
    "args": {
      "Python parent id": 12, "Python id": 3144, "finished": true, "Ev Idx": 6044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903764.728, "dur": 0.261,
    "args": {
      "Python parent id": 12, "Python id": 3145, "finished": true, "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903765.497, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 3146, "finished": true, "Ev Idx": 6046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903766.065, "dur": 9.840,
    "args": {
      "Python parent id": 12, "Python id": 3147, "finished": true, "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903777.256, "dur": 15.667,
    "args": {
      "Python parent id": 12, "Python id": 3148, "finished": true, "Ev Idx": 6048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903777.623, "dur": 12.064,
    "args": {
      "Python parent id": 3148, "Python id": 3149, "finished": true, "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903777.831, "dur": 0.488,
    "args": {
      "Python parent id": 3149, "Python id": 3150, "finished": true, "Ev Idx": 6050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903777.995, "dur": 0.244,
    "args": {
      "Python parent id": 3150, "Python id": 3151, "finished": true, "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903778.105, "dur": 0.100,
    "args": {
      "Python parent id": 3151, "Python id": 3152, "finished": true, "Ev Idx": 6052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903778.645, "dur": 8.976,
    "args": {
      "Python parent id": 3149, "Python id": 3153, "finished": true, "Ev Idx": 6053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903778.725, "dur": 0.071,
    "args": {
      "Python parent id": 3153, "Python id": 3154, "finished": true, "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903778.910, "dur": 0.063,
    "args": {
      "Python parent id": 3153, "Python id": 3155, "finished": true, "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903779.138, "dur": 0.065,
    "args": {
      "Python parent id": 3153, "Python id": 3156, "finished": true, "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903779.641, "dur": 0.065,
    "args": {
      "Python parent id": 3153, "Python id": 3157, "finished": true, "Ev Idx": 6057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903780.118, "dur": 0.077,
    "args": {
      "Python parent id": 3153, "Python id": 3158, "finished": true, "Ev Idx": 6058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903780.639, "dur": 6.915,
    "args": {
      "Python parent id": 3153, "Python id": 3159, "finished": true, "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903780.744, "dur": 0.065,
    "args": {
      "Python parent id": 3159, "Python id": 3160, "finished": true, "Ev Idx": 6060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903780.999, "dur": 0.055,
    "args": {
      "Python parent id": 3159, "Python id": 3161, "finished": true, "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903781.209, "dur": 0.069,
    "args": {
      "Python parent id": 3159, "Python id": 3162, "finished": true, "Ev Idx": 6062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903781.651, "dur": 0.030,
    "args": {
      "Python parent id": 3159, "Python id": 3163, "finished": true, "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.015, "dur": 5.456,
    "args": {
      "Python parent id": 3159, "Python id": 3164, "finished": true, "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.292, "dur": 5.119,
    "args": {
      "Python parent id": 3164, "Python id": 3165, "finished": true, "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.452, "dur": 3.215,
    "args": {
      "Python parent id": 3165, "Python id": 3166, "finished": true, "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.653, "dur": 2.934,
    "args": {
      "Python parent id": 3166, "Python id": 3167, "finished": true, "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.810, "dur": 0.290,
    "args": {
      "Python parent id": 3167, "Python id": 3168, "finished": true, "Ev Idx": 6068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903782.979, "dur": 0.063,
    "args": {
      "Python parent id": 3168, "Python id": 3169, "finished": true, "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903783.250, "dur": 2.013,
    "args": {
      "Python parent id": 3167, "Python id": 3170, "finished": true, "Ev Idx": 6070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903783.453, "dur": 1.733,
    "args": {
      "Python parent id": 3170, "Python id": 3171, "finished": true, "Ev Idx": 6071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903783.657, "dur": 1.462,
    "args": {
      "Python parent id": 3171, "Python id": 3172, "finished": true, "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903783.788, "dur": 1.140,
    "args": {
      "Python parent id": 3172, "Python id": 3173, "finished": true, "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903783.961, "dur": 0.499,
    "args": {
      "Python parent id": 3173, "Python id": 3174, "finished": true, "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903784.066, "dur": 0.051,
    "args": {
      "Python parent id": 3174, "Python id": 3175, "finished": true, "Ev Idx": 6075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903784.287, "dur": 0.114,
    "args": {
      "Python parent id": 3174, "Python id": 3176, "finished": true, "Ev Idx": 6076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903785.470, "dur": 0.066,
    "args": {
      "Python parent id": 3167, "Python id": 3177, "finished": true, "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903785.820, "dur": 0.079,
    "args": {
      "Python parent id": 3165, "Python id": 3178, "finished": true, "Ev Idx": 6078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903786.197, "dur": 1.173,
    "args": {
      "Python parent id": 3165, "Python id": 3179, "finished": true, "Ev Idx": 6079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903786.334, "dur": 0.992,
    "args": {
      "Python parent id": 3179, "Python id": 3180, "finished": true, "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903786.483, "dur": 0.400,
    "args": {
      "Python parent id": 3180, "Python id": 3181, "finished": true, "Ev Idx": 6081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903786.622, "dur": 0.200,
    "args": {
      "Python parent id": 3181, "Python id": 3182, "finished": true, "Ev Idx": 6082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903786.737, "dur": 0.054,
    "args": {
      "Python parent id": 3182, "Python id": 3183, "finished": true, "Ev Idx": 6083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903787.014, "dur": 0.282,
    "args": {
      "Python parent id": 3180, "Python id": 3184, "finished": true, "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903787.743, "dur": 0.200,
    "args": {
      "Python parent id": 3149, "Python id": 3185, "finished": true, "Ev Idx": 6085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903788.566, "dur": 0.968,
    "args": {
      "Python parent id": 3149, "Python id": 3186, "finished": true, "Ev Idx": 6086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903788.967, "dur": 0.516,
    "args": {
      "Python parent id": 3186, "Python id": 3187, "finished": true, "Ev Idx": 6087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903789.985, "dur": 2.889,
    "args": {
      "Python parent id": 3148, "Python id": 3188, "finished": true, "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903793.961, "dur": 92.317,
    "args": {
      "Python parent id": 12, "Python id": 3189, "finished": true, "Ev Idx": 6089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903794.343, "dur": 0.401,
    "args": {
      "Python parent id": 3189, "Python id": 3190, "finished": true, "Ev Idx": 6090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903794.609, "dur": 0.087,
    "args": {
      "Python parent id": 3190, "Python id": 3191, "finished": true, "Ev Idx": 6091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903795.086, "dur": 0.508,
    "args": {
      "Python parent id": 3189, "Python id": 3192, "finished": true, "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903795.807, "dur": 0.272,
    "args": {
      "Python parent id": 3189, "Python id": 3193, "finished": true, "Ev Idx": 6093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903796.318, "dur": 1.569,
    "args": {
      "Python parent id": 3189, "Python id": 3194, "finished": true, "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903798.447, "dur": 87.677,
    "args": {
      "Python parent id": 3189, "Python id": 3195, "finished": true, "Ev Idx": 6095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903799.347, "dur": 86.503,
    "args": {
      "Python parent id": 3195, "Python id": 3196, "finished": true, "Ev Idx": 6096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903799.621, "dur": 86.166,
    "args": {
      "Python parent id": 3196, "Python id": 3197, "finished": true, "Ev Idx": 6097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903805.304, "dur": 79.089,
    "args": {
      "Python parent id": 3197, "Python id": 3198, "finished": true, "Ev Idx": 6098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903805.639, "dur": 78.701,
    "args": {
      "Python parent id": 3198, "Python id": 3199, "finished": true, "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903806.269, "dur": 77.913,
    "args": {
      "Python parent id": 3199, "Python id": 3200, "finished": true, "Ev Idx": 6100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903806.871, "dur": 0.427,
    "args": {
      "Python parent id": 3200, "Python id": 3201, "finished": true, "Ev Idx": 6101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903807.148, "dur": 0.070,
    "args": {
      "Python parent id": 3201, "Python id": 3202, "finished": true, "Ev Idx": 6102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903807.650, "dur": 0.336,
    "args": {
      "Python parent id": 3200, "Python id": 3203, "finished": true, "Ev Idx": 6103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903807.881, "dur": 0.051,
    "args": {
      "Python parent id": 3203, "Python id": 3204, "finished": true, "Ev Idx": 6104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903808.382, "dur": 75.715,
    "args": {
      "Python parent id": 3200, "Python id": 3205, "finished": true, "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903809.034, "dur": 73.808,
    "args": {
      "Python parent id": 3205, "Python id": 3206, "finished": true, "Ev Idx": 6106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903816.600, "dur": 17.473,
    "args": {
      "Python parent id": 3206, "Python id": 3207, "finished": true, "Ev Idx": 6107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903816.740, "dur": 0.158,
    "args": {
      "Python parent id": 3207, "Python id": 3208, "finished": true, "Ev Idx": 6108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903817.213, "dur": 16.717,
    "args": {
      "Python parent id": 3207, "Python id": 3209, "finished": true, "Ev Idx": 6109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903817.401, "dur": 7.401,
    "args": {
      "Python parent id": 3209, "Python id": 3210, "finished": true, "Ev Idx": 6110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903842.035, "dur": 6.791,
    "args": {
      "Python parent id": 3206, "Python id": 3211, "finished": true, "Ev Idx": 6111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740903857.338, "dur": 8.587,
    "args": {
      "Python parent id": 3206, "Python id": 3212, "finished": true, "Ev Idx": 6112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903887.561, "dur": 17.533,
    "args": {
      "Python parent id": 12, "Python id": 3213, "finished": true, "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903888.024, "dur": 13.289,
    "args": {
      "Python parent id": 3213, "Python id": 3214, "finished": true, "Ev Idx": 6114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903888.180, "dur": 0.488,
    "args": {
      "Python parent id": 3214, "Python id": 3215, "finished": true, "Ev Idx": 6115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903888.360, "dur": 0.239,
    "args": {
      "Python parent id": 3215, "Python id": 3216, "finished": true, "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903888.485, "dur": 0.079,
    "args": {
      "Python parent id": 3216, "Python id": 3217, "finished": true, "Ev Idx": 6117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903888.997, "dur": 9.939,
    "args": {
      "Python parent id": 3214, "Python id": 3218, "finished": true, "Ev Idx": 6118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903889.077, "dur": 0.105,
    "args": {
      "Python parent id": 3218, "Python id": 3219, "finished": true, "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903889.306, "dur": 0.060,
    "args": {
      "Python parent id": 3218, "Python id": 3220, "finished": true, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903889.557, "dur": 0.056,
    "args": {
      "Python parent id": 3218, "Python id": 3221, "finished": true, "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903890.100, "dur": 0.029,
    "args": {
      "Python parent id": 3218, "Python id": 3222, "finished": true, "Ev Idx": 6122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903890.463, "dur": 0.076,
    "args": {
      "Python parent id": 3218, "Python id": 3223, "finished": true, "Ev Idx": 6123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903890.917, "dur": 7.954,
    "args": {
      "Python parent id": 3218, "Python id": 3224, "finished": true, "Ev Idx": 6124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903891.022, "dur": 0.065,
    "args": {
      "Python parent id": 3224, "Python id": 3225, "finished": true, "Ev Idx": 6125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903891.291, "dur": 0.062,
    "args": {
      "Python parent id": 3224, "Python id": 3226, "finished": true, "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903891.534, "dur": 0.065,
    "args": {
      "Python parent id": 3224, "Python id": 3227, "finished": true, "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903891.950, "dur": 0.037,
    "args": {
      "Python parent id": 3224, "Python id": 3228, "finished": true, "Ev Idx": 6128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903892.279, "dur": 6.505,
    "args": {
      "Python parent id": 3224, "Python id": 3229, "finished": true, "Ev Idx": 6129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903892.576, "dur": 6.131,
    "args": {
      "Python parent id": 3229, "Python id": 3230, "finished": true, "Ev Idx": 6130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903892.731, "dur": 4.031,
    "args": {
      "Python parent id": 3230, "Python id": 3231, "finished": true, "Ev Idx": 6131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903893.026, "dur": 3.657,
    "args": {
      "Python parent id": 3231, "Python id": 3232, "finished": true, "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903893.216, "dur": 0.572,
    "args": {
      "Python parent id": 3232, "Python id": 3233, "finished": true, "Ev Idx": 6133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903893.661, "dur": 0.058,
    "args": {
      "Python parent id": 3233, "Python id": 3234, "finished": true, "Ev Idx": 6134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903893.976, "dur": 2.310,
    "args": {
      "Python parent id": 3232, "Python id": 3235, "finished": true, "Ev Idx": 6135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903894.273, "dur": 1.939,
    "args": {
      "Python parent id": 3235, "Python id": 3236, "finished": true, "Ev Idx": 6136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903894.460, "dur": 1.671,
    "args": {
      "Python parent id": 3236, "Python id": 3237, "finished": true, "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903894.663, "dur": 1.233,
    "args": {
      "Python parent id": 3237, "Python id": 3238, "finished": true, "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903894.852, "dur": 0.519,
    "args": {
      "Python parent id": 3238, "Python id": 3239, "finished": true, "Ev Idx": 6139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903894.979, "dur": 0.045,
    "args": {
      "Python parent id": 3239, "Python id": 3240, "finished": true, "Ev Idx": 6140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903895.190, "dur": 0.125,
    "args": {
      "Python parent id": 3239, "Python id": 3241, "finished": true, "Ev Idx": 6141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903896.499, "dur": 0.120,
    "args": {
      "Python parent id": 3232, "Python id": 3242, "finished": true, "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903896.962, "dur": 0.112,
    "args": {
      "Python parent id": 3230, "Python id": 3243, "finished": true, "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903897.456, "dur": 1.186,
    "args": {
      "Python parent id": 3230, "Python id": 3244, "finished": true, "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903897.601, "dur": 0.994,
    "args": {
      "Python parent id": 3244, "Python id": 3245, "finished": true, "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903897.746, "dur": 0.385,
    "args": {
      "Python parent id": 3245, "Python id": 3246, "finished": true, "Ev Idx": 6146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903897.885, "dur": 0.186,
    "args": {
      "Python parent id": 3246, "Python id": 3247, "finished": true, "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903897.980, "dur": 0.060,
    "args": {
      "Python parent id": 3247, "Python id": 3248, "finished": true, "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903898.274, "dur": 0.290,
    "args": {
      "Python parent id": 3245, "Python id": 3249, "finished": true, "Ev Idx": 6149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903899.086, "dur": 0.227,
    "args": {
      "Python parent id": 3214, "Python id": 3250, "finished": true, "Ev Idx": 6150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903899.994, "dur": 1.133,
    "args": {
      "Python parent id": 3214, "Python id": 3251, "finished": true, "Ev Idx": 6151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903900.498, "dur": 0.553,
    "args": {
      "Python parent id": 3251, "Python id": 3252, "finished": true, "Ev Idx": 6152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903901.595, "dur": 3.421,
    "args": {
      "Python parent id": 3213, "Python id": 3253, "finished": true, "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903905.689, "dur": 7.326,
    "args": {
      "Python parent id": 12, "Python id": 3254, "finished": true, "Ev Idx": 6154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740903913.547, "dur": 0.131,
    "args": {
      "Python parent id": 12, "Python id": 3255, "finished": true, "Ev Idx": 6155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903914.296, "dur": 0.191,
    "args": {
      "Python parent id": 12, "Python id": 3256, "finished": true, "Ev Idx": 6156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903915.009, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 3257, "finished": true, "Ev Idx": 6157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740903915.491, "dur": 10.114,
    "args": {
      "Python parent id": 12, "Python id": 3258, "finished": true, "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740903927.099, "dur": 14.869,
    "args": {
      "Python parent id": 12, "Python id": 3259, "finished": true, "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740903927.409, "dur": 11.510,
    "args": {
      "Python parent id": 3259, "Python id": 3260, "finished": true, "Ev Idx": 6160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903927.556, "dur": 0.468,
    "args": {
      "Python parent id": 3260, "Python id": 3261, "finished": true, "Ev Idx": 6161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903927.704, "dur": 0.248,
    "args": {
      "Python parent id": 3261, "Python id": 3262, "finished": true, "Ev Idx": 6162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903927.806, "dur": 0.110,
    "args": {
      "Python parent id": 3262, "Python id": 3263, "finished": true, "Ev Idx": 6163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903928.319, "dur": 8.790,
    "args": {
      "Python parent id": 3260, "Python id": 3264, "finished": true, "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903928.408, "dur": 0.069,
    "args": {
      "Python parent id": 3264, "Python id": 3265, "finished": true, "Ev Idx": 6165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903928.618, "dur": 0.057,
    "args": {
      "Python parent id": 3264, "Python id": 3266, "finished": true, "Ev Idx": 6166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903928.832, "dur": 0.059,
    "args": {
      "Python parent id": 3264, "Python id": 3267, "finished": true, "Ev Idx": 6167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903929.315, "dur": 0.040,
    "args": {
      "Python parent id": 3264, "Python id": 3268, "finished": true, "Ev Idx": 6168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903929.670, "dur": 0.072,
    "args": {
      "Python parent id": 3264, "Python id": 3269, "finished": true, "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903930.047, "dur": 7.002,
    "args": {
      "Python parent id": 3264, "Python id": 3270, "finished": true, "Ev Idx": 6170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903930.143, "dur": 0.061,
    "args": {
      "Python parent id": 3270, "Python id": 3271, "finished": true, "Ev Idx": 6171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903930.348, "dur": 0.058,
    "args": {
      "Python parent id": 3270, "Python id": 3272, "finished": true, "Ev Idx": 6172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903930.586, "dur": 0.070,
    "args": {
      "Python parent id": 3270, "Python id": 3273, "finished": true, "Ev Idx": 6173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740903931.006, "dur": 0.038,
    "args": {
      "Python parent id": 3270, "Python id": 3274, "finished": true, "Ev Idx": 6174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740903931.413, "dur": 5.528,
    "args": {
      "Python parent id": 3270, "Python id": 3275, "finished": true, "Ev Idx": 6175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740903931.615, "dur": 5.265,
    "args": {
      "Python parent id": 3275, "Python id": 3276, "finished": true, "Ev Idx": 6176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740903931.781, "dur": 3.318,
    "args": {
      "Python parent id": 3276, "Python id": 3277, "finished": true, "Ev Idx": 6177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740903931.980, "dur": 3.043,
    "args": {
      "Python parent id": 3277, "Python id": 3278, "finished": true, "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740903932.145, "dur": 0.322,
    "args": {
      "Python parent id": 3278, "Python id": 3279, "finished": true, "Ev Idx": 6179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740903932.321, "dur": 0.068,
    "args": {
      "Python parent id": 3279, "Python id": 3280, "finished": true, "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740903932.621, "dur": 2.088,
    "args": {
      "Python parent id": 3278, "Python id": 3281, "finished": true, "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740903932.830, "dur": 1.791,
    "args": {
      "Python parent id": 3281, "Python id": 3282, "finished": true, "Ev Idx": 6182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740903933.025, "dur": 1.550,
    "args": {
      "Python parent id": 3282, "Python id": 3283, "finished": true, "Ev Idx": 6183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740903933.220, "dur": 1.161,
    "args": {
      "Python parent id": 3283, "Python id": 3284, "finished": true, "Ev Idx": 6184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740903933.405, "dur": 0.510,
    "args": {
      "Python parent id": 3284, "Python id": 3285, "finished": true, "Ev Idx": 6185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740903933.517, "dur": 0.049,
    "args": {
      "Python parent id": 3285, "Python id": 3286, "finished": true, "Ev Idx": 6186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740903933.741, "dur": 0.125,
    "args": {
      "Python parent id": 3285, "Python id": 3287, "finished": true, "Ev Idx": 6187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740903934.899, "dur": 0.068,
    "args": {
      "Python parent id": 3278, "Python id": 3288, "finished": true, "Ev Idx": 6188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740903935.246, "dur": 0.113,
    "args": {
      "Python parent id": 3276, "Python id": 3289, "finished": true, "Ev Idx": 6189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740903935.656, "dur": 1.175,
    "args": {
      "Python parent id": 3276, "Python id": 3290, "finished": true, "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740903935.819, "dur": 0.967,
    "args": {
      "Python parent id": 3290, "Python id": 3291, "finished": true, "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740903935.960, "dur": 0.395,
    "args": {
      "Python parent id": 3291, "Python id": 3292, "finished": true, "Ev Idx": 6192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740903936.107, "dur": 0.183,
    "args": {
      "Python parent id": 3292, "Python id": 3293, "finished": true, "Ev Idx": 6193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740903936.199, "dur": 0.059,
    "args": {
      "Python parent id": 3293, "Python id": 3294, "finished": true, "Ev Idx": 6194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740903936.488, "dur": 0.264,
    "args": {
      "Python parent id": 3291, "Python id": 3295, "finished": true, "Ev Idx": 6195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740903937.235, "dur": 0.189,
    "args": {
      "Python parent id": 3260, "Python id": 3296, "finished": true, "Ev Idx": 6196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740903937.973, "dur": 0.822,
    "args": {
      "Python parent id": 3260, "Python id": 3297, "finished": true, "Ev Idx": 6197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740903938.347, "dur": 0.398,
    "args": {
      "Python parent id": 3297, "Python id": 3298, "finished": true, "Ev Idx": 6198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740903939.201, "dur": 2.719,
    "args": {
      "Python parent id": 3259, "Python id": 3299, "finished": true, "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740903942.955, "dur": 106.837,
    "args": {
      "Python parent id": 12, "Python id": 3300, "finished": true, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903943.258, "dur": 0.449,
    "args": {
      "Python parent id": 3300, "Python id": 3301, "finished": true, "Ev Idx": 6201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903943.542, "dur": 0.080,
    "args": {
      "Python parent id": 3301, "Python id": 3302, "finished": true, "Ev Idx": 6202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903944.036, "dur": 0.365,
    "args": {
      "Python parent id": 3300, "Python id": 3303, "finished": true, "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740903944.599, "dur": 0.304,
    "args": {
      "Python parent id": 3300, "Python id": 3304, "finished": true, "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740903945.163, "dur": 1.701,
    "args": {
      "Python parent id": 3300, "Python id": 3305, "finished": true, "Ev Idx": 6205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740903947.406, "dur": 102.298,
    "args": {
      "Python parent id": 3300, "Python id": 3306, "finished": true, "Ev Idx": 6206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740903948.313, "dur": 101.145,
    "args": {
      "Python parent id": 3306, "Python id": 3307, "finished": true, "Ev Idx": 6207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903948.629, "dur": 100.751,
    "args": {
      "Python parent id": 3307, "Python id": 3308, "finished": true, "Ev Idx": 6208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740903953.943, "dur": 93.824,
    "args": {
      "Python parent id": 3308, "Python id": 3309, "finished": true, "Ev Idx": 6209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740903954.324, "dur": 93.376,
    "args": {
      "Python parent id": 3309, "Python id": 3310, "finished": true, "Ev Idx": 6210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740903954.893, "dur": 92.634,
    "args": {
      "Python parent id": 3310, "Python id": 3311, "finished": true, "Ev Idx": 6211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903955.417, "dur": 0.429,
    "args": {
      "Python parent id": 3311, "Python id": 3312, "finished": true, "Ev Idx": 6212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903955.703, "dur": 0.058,
    "args": {
      "Python parent id": 3312, "Python id": 3313, "finished": true, "Ev Idx": 6213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740903956.198, "dur": 0.322,
    "args": {
      "Python parent id": 3311, "Python id": 3314, "finished": true, "Ev Idx": 6214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740903956.411, "dur": 0.052,
    "args": {
      "Python parent id": 3314, "Python id": 3315, "finished": true, "Ev Idx": 6215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903956.859, "dur": 90.528,
    "args": {
      "Python parent id": 3311, "Python id": 3316, "finished": true, "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740903957.489, "dur": 88.459,
    "args": {
      "Python parent id": 3316, "Python id": 3317, "finished": true, "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740903964.845, "dur": 34.061,
    "args": {
      "Python parent id": 3317, "Python id": 3318, "finished": true, "Ev Idx": 6218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740903964.997, "dur": 0.142,
    "args": {
      "Python parent id": 3318, "Python id": 3319, "finished": true, "Ev Idx": 6219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740903965.449, "dur": 33.325,
    "args": {
      "Python parent id": 3318, "Python id": 3320, "finished": true, "Ev Idx": 6220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740903965.572, "dur": 23.684,
    "args": {
      "Python parent id": 3320, "Python id": 3321, "finished": true, "Ev Idx": 6221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904006.528, "dur": 7.451,
    "args": {
      "Python parent id": 3317, "Python id": 3322, "finished": true, "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904022.505, "dur": 8.320,
    "args": {
      "Python parent id": 3317, "Python id": 3323, "finished": true, "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904050.982, "dur": 16.837,
    "args": {
      "Python parent id": 12, "Python id": 3324, "finished": true, "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904051.413, "dur": 12.686,
    "args": {
      "Python parent id": 3324, "Python id": 3325, "finished": true, "Ev Idx": 6225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904051.589, "dur": 0.528,
    "args": {
      "Python parent id": 3325, "Python id": 3326, "finished": true, "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904051.825, "dur": 0.230,
    "args": {
      "Python parent id": 3326, "Python id": 3327, "finished": true, "Ev Idx": 6227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904051.941, "dur": 0.081,
    "args": {
      "Python parent id": 3327, "Python id": 3328, "finished": true, "Ev Idx": 6228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904052.497, "dur": 9.333,
    "args": {
      "Python parent id": 3325, "Python id": 3329, "finished": true, "Ev Idx": 6229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904052.636, "dur": 0.076,
    "args": {
      "Python parent id": 3329, "Python id": 3330, "finished": true, "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904052.876, "dur": 0.049,
    "args": {
      "Python parent id": 3329, "Python id": 3331, "finished": true, "Ev Idx": 6231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904053.064, "dur": 0.059,
    "args": {
      "Python parent id": 3329, "Python id": 3332, "finished": true, "Ev Idx": 6232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904053.665, "dur": 0.040,
    "args": {
      "Python parent id": 3329, "Python id": 3333, "finished": true, "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904054.002, "dur": 0.075,
    "args": {
      "Python parent id": 3329, "Python id": 3334, "finished": true, "Ev Idx": 6234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904054.494, "dur": 7.279,
    "args": {
      "Python parent id": 3329, "Python id": 3335, "finished": true, "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904054.593, "dur": 0.066,
    "args": {
      "Python parent id": 3335, "Python id": 3336, "finished": true, "Ev Idx": 6236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904054.816, "dur": 0.066,
    "args": {
      "Python parent id": 3335, "Python id": 3337, "finished": true, "Ev Idx": 6237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904055.060, "dur": 0.070,
    "args": {
      "Python parent id": 3335, "Python id": 3338, "finished": true, "Ev Idx": 6238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904055.444, "dur": 0.036,
    "args": {
      "Python parent id": 3335, "Python id": 3339, "finished": true, "Ev Idx": 6239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904055.750, "dur": 5.911,
    "args": {
      "Python parent id": 3335, "Python id": 3340, "finished": true, "Ev Idx": 6240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904056.061, "dur": 5.541,
    "args": {
      "Python parent id": 3340, "Python id": 3341, "finished": true, "Ev Idx": 6241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904056.185, "dur": 3.653,
    "args": {
      "Python parent id": 3341, "Python id": 3342, "finished": true, "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904056.438, "dur": 3.318,
    "args": {
      "Python parent id": 3342, "Python id": 3343, "finished": true, "Ev Idx": 6243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904056.587, "dur": 0.291,
    "args": {
      "Python parent id": 3343, "Python id": 3344, "finished": true, "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904056.752, "dur": 0.071,
    "args": {
      "Python parent id": 3344, "Python id": 3345, "finished": true, "Ev Idx": 6245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904057.046, "dur": 2.316,
    "args": {
      "Python parent id": 3343, "Python id": 3346, "finished": true, "Ev Idx": 6246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904057.300, "dur": 1.976,
    "args": {
      "Python parent id": 3346, "Python id": 3347, "finished": true, "Ev Idx": 6247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904057.571, "dur": 1.657,
    "args": {
      "Python parent id": 3347, "Python id": 3348, "finished": true, "Ev Idx": 6248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904057.776, "dur": 1.206,
    "args": {
      "Python parent id": 3348, "Python id": 3349, "finished": true, "Ev Idx": 6249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904057.941, "dur": 0.471,
    "args": {
      "Python parent id": 3349, "Python id": 3350, "finished": true, "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904058.028, "dur": 0.050,
    "args": {
      "Python parent id": 3350, "Python id": 3351, "finished": true, "Ev Idx": 6251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904058.250, "dur": 0.123,
    "args": {
      "Python parent id": 3350, "Python id": 3352, "finished": true, "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904059.595, "dur": 0.124,
    "args": {
      "Python parent id": 3343, "Python id": 3353, "finished": true, "Ev Idx": 6253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904059.987, "dur": 0.117,
    "args": {
      "Python parent id": 3341, "Python id": 3354, "finished": true, "Ev Idx": 6254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904060.405, "dur": 1.154,
    "args": {
      "Python parent id": 3341, "Python id": 3355, "finished": true, "Ev Idx": 6255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904060.541, "dur": 0.955,
    "args": {
      "Python parent id": 3355, "Python id": 3356, "finished": true, "Ev Idx": 6256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904060.708, "dur": 0.382,
    "args": {
      "Python parent id": 3356, "Python id": 3357, "finished": true, "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904060.845, "dur": 0.186,
    "args": {
      "Python parent id": 3357, "Python id": 3358, "finished": true, "Ev Idx": 6258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904060.936, "dur": 0.062,
    "args": {
      "Python parent id": 3358, "Python id": 3359, "finished": true, "Ev Idx": 6259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904061.224, "dur": 0.230,
    "args": {
      "Python parent id": 3356, "Python id": 3360, "finished": true, "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904061.956, "dur": 0.259,
    "args": {
      "Python parent id": 3325, "Python id": 3361, "finished": true, "Ev Idx": 6261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904062.874, "dur": 1.068,
    "args": {
      "Python parent id": 3325, "Python id": 3362, "finished": true, "Ev Idx": 6262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904063.400, "dur": 0.492,
    "args": {
      "Python parent id": 3362, "Python id": 3363, "finished": true, "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904064.401, "dur": 3.371,
    "args": {
      "Python parent id": 3324, "Python id": 3364, "finished": true, "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904068.350, "dur": 9.127,
    "args": {
      "Python parent id": 12, "Python id": 3365, "finished": true, "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904077.980, "dur": 0.115,
    "args": {
      "Python parent id": 12, "Python id": 3366, "finished": true, "Ev Idx": 6266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904078.709, "dur": 0.188,
    "args": {
      "Python parent id": 12, "Python id": 3367, "finished": true, "Ev Idx": 6267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904079.373, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 3368, "finished": true, "Ev Idx": 6268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904079.883, "dur": 10.999,
    "args": {
      "Python parent id": 12, "Python id": 3369, "finished": true, "Ev Idx": 6269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904092.219, "dur": 15.222,
    "args": {
      "Python parent id": 12, "Python id": 3370, "finished": true, "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904092.554, "dur": 11.645,
    "args": {
      "Python parent id": 3370, "Python id": 3371, "finished": true, "Ev Idx": 6271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904092.767, "dur": 0.454,
    "args": {
      "Python parent id": 3371, "Python id": 3372, "finished": true, "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904092.912, "dur": 0.235,
    "args": {
      "Python parent id": 3372, "Python id": 3373, "finished": true, "Ev Idx": 6273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904093.015, "dur": 0.096,
    "args": {
      "Python parent id": 3373, "Python id": 3374, "finished": true, "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904093.604, "dur": 8.638,
    "args": {
      "Python parent id": 3371, "Python id": 3375, "finished": true, "Ev Idx": 6275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904093.697, "dur": 0.070,
    "args": {
      "Python parent id": 3375, "Python id": 3376, "finished": true, "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904093.879, "dur": 0.063,
    "args": {
      "Python parent id": 3375, "Python id": 3377, "finished": true, "Ev Idx": 6277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904094.126, "dur": 0.059,
    "args": {
      "Python parent id": 3375, "Python id": 3378, "finished": true, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904094.593, "dur": 0.041,
    "args": {
      "Python parent id": 3375, "Python id": 3379, "finished": true, "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904094.910, "dur": 0.076,
    "args": {
      "Python parent id": 3375, "Python id": 3380, "finished": true, "Ev Idx": 6280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904095.311, "dur": 6.871,
    "args": {
      "Python parent id": 3375, "Python id": 3381, "finished": true, "Ev Idx": 6281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904095.409, "dur": 0.064,
    "args": {
      "Python parent id": 3381, "Python id": 3382, "finished": true, "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904095.629, "dur": 0.062,
    "args": {
      "Python parent id": 3381, "Python id": 3383, "finished": true, "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904095.859, "dur": 0.071,
    "args": {
      "Python parent id": 3381, "Python id": 3384, "finished": true, "Ev Idx": 6284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904096.320, "dur": 0.038,
    "args": {
      "Python parent id": 3381, "Python id": 3385, "finished": true, "Ev Idx": 6285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904096.689, "dur": 5.399,
    "args": {
      "Python parent id": 3381, "Python id": 3386, "finished": true, "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904096.982, "dur": 5.045,
    "args": {
      "Python parent id": 3386, "Python id": 3387, "finished": true, "Ev Idx": 6287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904097.109, "dur": 3.069,
    "args": {
      "Python parent id": 3387, "Python id": 3388, "finished": true, "Ev Idx": 6288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904097.293, "dur": 2.842,
    "args": {
      "Python parent id": 3388, "Python id": 3389, "finished": true, "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904097.477, "dur": 0.257,
    "args": {
      "Python parent id": 3389, "Python id": 3390, "finished": true, "Ev Idx": 6290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904097.590, "dur": 0.075,
    "args": {
      "Python parent id": 3390, "Python id": 3391, "finished": true, "Ev Idx": 6291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904097.909, "dur": 1.900,
    "args": {
      "Python parent id": 3389, "Python id": 3392, "finished": true, "Ev Idx": 6292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.102, "dur": 1.650,
    "args": {
      "Python parent id": 3392, "Python id": 3393, "finished": true, "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.290, "dur": 1.411,
    "args": {
      "Python parent id": 3393, "Python id": 3394, "finished": true, "Ev Idx": 6294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.435, "dur": 1.099,
    "args": {
      "Python parent id": 3394, "Python id": 3395, "finished": true, "Ev Idx": 6295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.592, "dur": 0.439,
    "args": {
      "Python parent id": 3395, "Python id": 3396, "finished": true, "Ev Idx": 6296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.678, "dur": 0.051,
    "args": {
      "Python parent id": 3396, "Python id": 3397, "finished": true, "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904098.862, "dur": 0.125,
    "args": {
      "Python parent id": 3396, "Python id": 3398, "finished": true, "Ev Idx": 6298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904100.018, "dur": 0.061,
    "args": {
      "Python parent id": 3389, "Python id": 3399, "finished": true, "Ev Idx": 6299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904100.358, "dur": 0.109,
    "args": {
      "Python parent id": 3387, "Python id": 3400, "finished": true, "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904100.774, "dur": 1.206,
    "args": {
      "Python parent id": 3387, "Python id": 3401, "finished": true, "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904100.919, "dur": 1.012,
    "args": {
      "Python parent id": 3401, "Python id": 3402, "finished": true, "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904101.048, "dur": 0.393,
    "args": {
      "Python parent id": 3402, "Python id": 3403, "finished": true, "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904101.188, "dur": 0.190,
    "args": {
      "Python parent id": 3403, "Python id": 3404, "finished": true, "Ev Idx": 6304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904101.282, "dur": 0.059,
    "args": {
      "Python parent id": 3404, "Python id": 3405, "finished": true, "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904101.579, "dur": 0.320,
    "args": {
      "Python parent id": 3402, "Python id": 3406, "finished": true, "Ev Idx": 6306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904102.367, "dur": 0.219,
    "args": {
      "Python parent id": 3371, "Python id": 3407, "finished": true, "Ev Idx": 6307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904103.221, "dur": 0.838,
    "args": {
      "Python parent id": 3371, "Python id": 3408, "finished": true, "Ev Idx": 6308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904103.576, "dur": 0.434,
    "args": {
      "Python parent id": 3408, "Python id": 3409, "finished": true, "Ev Idx": 6309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904104.466, "dur": 2.927,
    "args": {
      "Python parent id": 3370, "Python id": 3410, "finished": true, "Ev Idx": 6310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904108.498, "dur": 94.540,
    "args": {
      "Python parent id": 12, "Python id": 3411, "finished": true, "Ev Idx": 6311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904108.874, "dur": 0.423,
    "args": {
      "Python parent id": 3411, "Python id": 3412, "finished": true, "Ev Idx": 6312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904109.153, "dur": 0.070,
    "args": {
      "Python parent id": 3412, "Python id": 3413, "finished": true, "Ev Idx": 6313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904109.582, "dur": 0.437,
    "args": {
      "Python parent id": 3411, "Python id": 3414, "finished": true, "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904110.263, "dur": 0.287,
    "args": {
      "Python parent id": 3411, "Python id": 3415, "finished": true, "Ev Idx": 6315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904110.780, "dur": 1.663,
    "args": {
      "Python parent id": 3411, "Python id": 3416, "finished": true, "Ev Idx": 6316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904113.050, "dur": 89.873,
    "args": {
      "Python parent id": 3411, "Python id": 3417, "finished": true, "Ev Idx": 6317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904113.981, "dur": 88.705,
    "args": {
      "Python parent id": 3417, "Python id": 3418, "finished": true, "Ev Idx": 6318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904114.262, "dur": 88.338,
    "args": {
      "Python parent id": 3418, "Python id": 3419, "finished": true, "Ev Idx": 6319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904120.372, "dur": 80.859,
    "args": {
      "Python parent id": 3419, "Python id": 3420, "finished": true, "Ev Idx": 6320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904120.678, "dur": 80.467,
    "args": {
      "Python parent id": 3420, "Python id": 3421, "finished": true, "Ev Idx": 6321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904121.232, "dur": 79.681,
    "args": {
      "Python parent id": 3421, "Python id": 3422, "finished": true, "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904121.841, "dur": 0.417,
    "args": {
      "Python parent id": 3422, "Python id": 3423, "finished": true, "Ev Idx": 6323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904122.127, "dur": 0.060,
    "args": {
      "Python parent id": 3423, "Python id": 3424, "finished": true, "Ev Idx": 6324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904122.565, "dur": 0.308,
    "args": {
      "Python parent id": 3422, "Python id": 3425, "finished": true, "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904122.774, "dur": 0.047,
    "args": {
      "Python parent id": 3425, "Python id": 3426, "finished": true, "Ev Idx": 6326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904123.237, "dur": 77.593,
    "args": {
      "Python parent id": 3422, "Python id": 3427, "finished": true, "Ev Idx": 6327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904123.963, "dur": 75.424,
    "args": {
      "Python parent id": 3427, "Python id": 3428, "finished": true, "Ev Idx": 6328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904132.059, "dur": 18.638,
    "args": {
      "Python parent id": 3428, "Python id": 3429, "finished": true, "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904132.175, "dur": 0.151,
    "args": {
      "Python parent id": 3429, "Python id": 3430, "finished": true, "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904132.657, "dur": 17.898,
    "args": {
      "Python parent id": 3429, "Python id": 3431, "finished": true, "Ev Idx": 6331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904132.842, "dur": 7.780,
    "args": {
      "Python parent id": 3431, "Python id": 3432, "finished": true, "Ev Idx": 6332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904159.031, "dur": 7.560,
    "args": {
      "Python parent id": 3428, "Python id": 3433, "finished": true, "Ev Idx": 6333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904175.523, "dur": 9.369,
    "args": {
      "Python parent id": 3428, "Python id": 3434, "finished": true, "Ev Idx": 6334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904204.380, "dur": 19.440,
    "args": {
      "Python parent id": 12, "Python id": 3435, "finished": true, "Ev Idx": 6335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904204.783, "dur": 14.609,
    "args": {
      "Python parent id": 3435, "Python id": 3436, "finished": true, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904204.946, "dur": 0.568,
    "args": {
      "Python parent id": 3436, "Python id": 3437, "finished": true, "Ev Idx": 6337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904205.170, "dur": 0.277,
    "args": {
      "Python parent id": 3437, "Python id": 3438, "finished": true, "Ev Idx": 6338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904205.274, "dur": 0.117,
    "args": {
      "Python parent id": 3438, "Python id": 3439, "finished": true, "Ev Idx": 6339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904205.868, "dur": 11.090,
    "args": {
      "Python parent id": 3436, "Python id": 3440, "finished": true, "Ev Idx": 6340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904206.025, "dur": 0.076,
    "args": {
      "Python parent id": 3440, "Python id": 3441, "finished": true, "Ev Idx": 6341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904206.192, "dur": 0.062,
    "args": {
      "Python parent id": 3440, "Python id": 3442, "finished": true, "Ev Idx": 6342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904206.423, "dur": 0.065,
    "args": {
      "Python parent id": 3440, "Python id": 3443, "finished": true, "Ev Idx": 6343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904206.993, "dur": 0.067,
    "args": {
      "Python parent id": 3440, "Python id": 3444, "finished": true, "Ev Idx": 6344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904207.358, "dur": 0.095,
    "args": {
      "Python parent id": 3440, "Python id": 3445, "finished": true, "Ev Idx": 6345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904207.908, "dur": 8.950,
    "args": {
      "Python parent id": 3440, "Python id": 3446, "finished": true, "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904208.012, "dur": 0.066,
    "args": {
      "Python parent id": 3446, "Python id": 3447, "finished": true, "Ev Idx": 6347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904208.231, "dur": 0.061,
    "args": {
      "Python parent id": 3446, "Python id": 3448, "finished": true, "Ev Idx": 6348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904208.477, "dur": 0.070,
    "args": {
      "Python parent id": 3446, "Python id": 3449, "finished": true, "Ev Idx": 6349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904208.907, "dur": 0.029,
    "args": {
      "Python parent id": 3446, "Python id": 3450, "finished": true, "Ev Idx": 6350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904209.275, "dur": 7.523,
    "args": {
      "Python parent id": 3446, "Python id": 3451, "finished": true, "Ev Idx": 6351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904209.567, "dur": 7.171,
    "args": {
      "Python parent id": 3451, "Python id": 3452, "finished": true, "Ev Idx": 6352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904209.704, "dur": 3.681,
    "args": {
      "Python parent id": 3452, "Python id": 3453, "finished": true, "Ev Idx": 6353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904209.899, "dur": 3.437,
    "args": {
      "Python parent id": 3453, "Python id": 3454, "finished": true, "Ev Idx": 6354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904210.087, "dur": 0.273,
    "args": {
      "Python parent id": 3454, "Python id": 3455, "finished": true, "Ev Idx": 6355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904210.203, "dur": 0.073,
    "args": {
      "Python parent id": 3455, "Python id": 3456, "finished": true, "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904210.568, "dur": 2.334,
    "args": {
      "Python parent id": 3454, "Python id": 3457, "finished": true, "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904210.862, "dur": 1.947,
    "args": {
      "Python parent id": 3457, "Python id": 3458, "finished": true, "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904211.110, "dur": 1.632,
    "args": {
      "Python parent id": 3458, "Python id": 3459, "finished": true, "Ev Idx": 6359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904211.250, "dur": 1.269,
    "args": {
      "Python parent id": 3459, "Python id": 3460, "finished": true, "Ev Idx": 6360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904211.463, "dur": 0.546,
    "args": {
      "Python parent id": 3460, "Python id": 3461, "finished": true, "Ev Idx": 6361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904211.603, "dur": 0.041,
    "args": {
      "Python parent id": 3461, "Python id": 3462, "finished": true, "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904211.842, "dur": 0.124,
    "args": {
      "Python parent id": 3461, "Python id": 3463, "finished": true, "Ev Idx": 6363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904213.162, "dur": 0.116,
    "args": {
      "Python parent id": 3454, "Python id": 3464, "finished": true, "Ev Idx": 6364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904214.936, "dur": 0.131,
    "args": {
      "Python parent id": 3452, "Python id": 3465, "finished": true, "Ev Idx": 6365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904215.382, "dur": 1.281,
    "args": {
      "Python parent id": 3452, "Python id": 3466, "finished": true, "Ev Idx": 6366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904215.555, "dur": 1.048,
    "args": {
      "Python parent id": 3466, "Python id": 3467, "finished": true, "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904215.718, "dur": 0.412,
    "args": {
      "Python parent id": 3467, "Python id": 3468, "finished": true, "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904215.866, "dur": 0.182,
    "args": {
      "Python parent id": 3468, "Python id": 3469, "finished": true, "Ev Idx": 6369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904215.960, "dur": 0.056,
    "args": {
      "Python parent id": 3469, "Python id": 3470, "finished": true, "Ev Idx": 6370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904216.266, "dur": 0.294,
    "args": {
      "Python parent id": 3467, "Python id": 3471, "finished": true, "Ev Idx": 6371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904217.108, "dur": 0.251,
    "args": {
      "Python parent id": 3436, "Python id": 3472, "finished": true, "Ev Idx": 6372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904218.083, "dur": 1.157,
    "args": {
      "Python parent id": 3436, "Python id": 3473, "finished": true, "Ev Idx": 6373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904218.665, "dur": 0.523,
    "args": {
      "Python parent id": 3473, "Python id": 3474, "finished": true, "Ev Idx": 6374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904219.751, "dur": 3.988,
    "args": {
      "Python parent id": 3435, "Python id": 3475, "finished": true, "Ev Idx": 6375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904224.418, "dur": 7.893,
    "args": {
      "Python parent id": 12, "Python id": 3476, "finished": true, "Ev Idx": 6376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904232.788, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 3477, "finished": true, "Ev Idx": 6377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904233.520, "dur": 0.212,
    "args": {
      "Python parent id": 12, "Python id": 3478, "finished": true, "Ev Idx": 6378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904234.221, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 3479, "finished": true, "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904234.725, "dur": 10.051,
    "args": {
      "Python parent id": 12, "Python id": 3480, "finished": true, "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904246.155, "dur": 15.278,
    "args": {
      "Python parent id": 12, "Python id": 3481, "finished": true, "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904246.477, "dur": 11.738,
    "args": {
      "Python parent id": 3481, "Python id": 3482, "finished": true, "Ev Idx": 6382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904246.633, "dur": 0.490,
    "args": {
      "Python parent id": 3482, "Python id": 3483, "finished": true, "Ev Idx": 6383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904246.789, "dur": 0.236,
    "args": {
      "Python parent id": 3483, "Python id": 3484, "finished": true, "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904246.889, "dur": 0.100,
    "args": {
      "Python parent id": 3484, "Python id": 3485, "finished": true, "Ev Idx": 6385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904247.480, "dur": 8.854,
    "args": {
      "Python parent id": 3482, "Python id": 3486, "finished": true, "Ev Idx": 6386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904247.637, "dur": 0.070,
    "args": {
      "Python parent id": 3486, "Python id": 3487, "finished": true, "Ev Idx": 6387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904247.827, "dur": 0.068,
    "args": {
      "Python parent id": 3486, "Python id": 3488, "finished": true, "Ev Idx": 6388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904248.064, "dur": 0.062,
    "args": {
      "Python parent id": 3486, "Python id": 3489, "finished": true, "Ev Idx": 6389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904248.619, "dur": 0.068,
    "args": {
      "Python parent id": 3486, "Python id": 3490, "finished": true, "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904248.995, "dur": 0.082,
    "args": {
      "Python parent id": 3486, "Python id": 3491, "finished": true, "Ev Idx": 6391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904249.447, "dur": 6.804,
    "args": {
      "Python parent id": 3486, "Python id": 3492, "finished": true, "Ev Idx": 6392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904249.532, "dur": 0.068,
    "args": {
      "Python parent id": 3492, "Python id": 3493, "finished": true, "Ev Idx": 6393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904249.758, "dur": 0.059,
    "args": {
      "Python parent id": 3492, "Python id": 3494, "finished": true, "Ev Idx": 6394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904249.996, "dur": 0.076,
    "args": {
      "Python parent id": 3492, "Python id": 3495, "finished": true, "Ev Idx": 6395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904250.396, "dur": 0.047,
    "args": {
      "Python parent id": 3492, "Python id": 3496, "finished": true, "Ev Idx": 6396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904250.793, "dur": 5.396,
    "args": {
      "Python parent id": 3492, "Python id": 3497, "finished": true, "Ev Idx": 6397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904251.021, "dur": 5.108,
    "args": {
      "Python parent id": 3497, "Python id": 3498, "finished": true, "Ev Idx": 6398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904251.159, "dur": 3.279,
    "args": {
      "Python parent id": 3498, "Python id": 3499, "finished": true, "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904251.352, "dur": 3.041,
    "args": {
      "Python parent id": 3499, "Python id": 3500, "finished": true, "Ev Idx": 6400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904251.505, "dur": 0.355,
    "args": {
      "Python parent id": 3500, "Python id": 3501, "finished": true, "Ev Idx": 6401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904251.704, "dur": 0.072,
    "args": {
      "Python parent id": 3501, "Python id": 3502, "finished": true, "Ev Idx": 6402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.034, "dur": 2.003,
    "args": {
      "Python parent id": 3500, "Python id": 3503, "finished": true, "Ev Idx": 6403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.220, "dur": 1.740,
    "args": {
      "Python parent id": 3503, "Python id": 3504, "finished": true, "Ev Idx": 6404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.451, "dur": 1.462,
    "args": {
      "Python parent id": 3504, "Python id": 3505, "finished": true, "Ev Idx": 6405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.601, "dur": 1.104,
    "args": {
      "Python parent id": 3505, "Python id": 3506, "finished": true, "Ev Idx": 6406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.769, "dur": 0.479,
    "args": {
      "Python parent id": 3506, "Python id": 3507, "finished": true, "Ev Idx": 6407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904252.880, "dur": 0.054,
    "args": {
      "Python parent id": 3507, "Python id": 3508, "finished": true, "Ev Idx": 6408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904253.071, "dur": 0.121,
    "args": {
      "Python parent id": 3507, "Python id": 3509, "finished": true, "Ev Idx": 6409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904254.262, "dur": 0.074,
    "args": {
      "Python parent id": 3500, "Python id": 3510, "finished": true, "Ev Idx": 6410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904254.607, "dur": 0.094,
    "args": {
      "Python parent id": 3498, "Python id": 3511, "finished": true, "Ev Idx": 6411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904254.945, "dur": 1.141,
    "args": {
      "Python parent id": 3498, "Python id": 3512, "finished": true, "Ev Idx": 6412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904255.080, "dur": 0.962,
    "args": {
      "Python parent id": 3512, "Python id": 3513, "finished": true, "Ev Idx": 6413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904255.272, "dur": 0.351,
    "args": {
      "Python parent id": 3513, "Python id": 3514, "finished": true, "Ev Idx": 6414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904255.400, "dur": 0.175,
    "args": {
      "Python parent id": 3514, "Python id": 3515, "finished": true, "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904255.485, "dur": 0.059,
    "args": {
      "Python parent id": 3515, "Python id": 3516, "finished": true, "Ev Idx": 6416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904255.751, "dur": 0.257,
    "args": {
      "Python parent id": 3513, "Python id": 3517, "finished": true, "Ev Idx": 6417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904256.489, "dur": 0.176,
    "args": {
      "Python parent id": 3482, "Python id": 3518, "finished": true, "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904257.259, "dur": 0.818,
    "args": {
      "Python parent id": 3482, "Python id": 3519, "finished": true, "Ev Idx": 6419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904257.628, "dur": 0.400,
    "args": {
      "Python parent id": 3519, "Python id": 3520, "finished": true, "Ev Idx": 6420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904258.515, "dur": 2.873,
    "args": {
      "Python parent id": 3481, "Python id": 3521, "finished": true, "Ev Idx": 6421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904262.415, "dur": 88.765,
    "args": {
      "Python parent id": 12, "Python id": 3522, "finished": true, "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904262.759, "dur": 0.424,
    "args": {
      "Python parent id": 3522, "Python id": 3523, "finished": true, "Ev Idx": 6423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904263.055, "dur": 0.073,
    "args": {
      "Python parent id": 3523, "Python id": 3524, "finished": true, "Ev Idx": 6424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904263.517, "dur": 0.418,
    "args": {
      "Python parent id": 3522, "Python id": 3525, "finished": true, "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904264.152, "dur": 0.292,
    "args": {
      "Python parent id": 3522, "Python id": 3526, "finished": true, "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904264.708, "dur": 1.699,
    "args": {
      "Python parent id": 3522, "Python id": 3527, "finished": true, "Ev Idx": 6427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904266.984, "dur": 84.097,
    "args": {
      "Python parent id": 3522, "Python id": 3528, "finished": true, "Ev Idx": 6428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904267.780, "dur": 83.031,
    "args": {
      "Python parent id": 3528, "Python id": 3529, "finished": true, "Ev Idx": 6429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904268.077, "dur": 82.641,
    "args": {
      "Python parent id": 3529, "Python id": 3530, "finished": true, "Ev Idx": 6430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904273.708, "dur": 75.629,
    "args": {
      "Python parent id": 3530, "Python id": 3531, "finished": true, "Ev Idx": 6431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904273.988, "dur": 75.292,
    "args": {
      "Python parent id": 3531, "Python id": 3532, "finished": true, "Ev Idx": 6432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904274.601, "dur": 74.496,
    "args": {
      "Python parent id": 3532, "Python id": 3533, "finished": true, "Ev Idx": 6433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904275.168, "dur": 0.464,
    "args": {
      "Python parent id": 3533, "Python id": 3534, "finished": true, "Ev Idx": 6434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904275.447, "dur": 0.063,
    "args": {
      "Python parent id": 3534, "Python id": 3535, "finished": true, "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904275.961, "dur": 0.338,
    "args": {
      "Python parent id": 3533, "Python id": 3536, "finished": true, "Ev Idx": 6436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904276.198, "dur": 0.047,
    "args": {
      "Python parent id": 3536, "Python id": 3537, "finished": true, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904276.645, "dur": 72.367,
    "args": {
      "Python parent id": 3533, "Python id": 3538, "finished": true, "Ev Idx": 6438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904277.324, "dur": 70.329,
    "args": {
      "Python parent id": 3538, "Python id": 3539, "finished": true, "Ev Idx": 6439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904284.727, "dur": 17.366,
    "args": {
      "Python parent id": 3539, "Python id": 3540, "finished": true, "Ev Idx": 6440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904284.854, "dur": 0.151,
    "args": {
      "Python parent id": 3540, "Python id": 3541, "finished": true, "Ev Idx": 6441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904285.332, "dur": 16.623,
    "args": {
      "Python parent id": 3540, "Python id": 3542, "finished": true, "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904285.441, "dur": 7.176,
    "args": {
      "Python parent id": 3542, "Python id": 3543, "finished": true, "Ev Idx": 6443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904309.788, "dur": 6.971,
    "args": {
      "Python parent id": 3539, "Python id": 3544, "finished": true, "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904325.173, "dur": 7.865,
    "args": {
      "Python parent id": 3539, "Python id": 3545, "finished": true, "Ev Idx": 6445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904352.384, "dur": 18.786,
    "args": {
      "Python parent id": 12, "Python id": 3546, "finished": true, "Ev Idx": 6446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904352.751, "dur": 13.450,
    "args": {
      "Python parent id": 3546, "Python id": 3547, "finished": true, "Ev Idx": 6447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904352.962, "dur": 0.474,
    "args": {
      "Python parent id": 3547, "Python id": 3548, "finished": true, "Ev Idx": 6448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904353.156, "dur": 0.212,
    "args": {
      "Python parent id": 3548, "Python id": 3549, "finished": true, "Ev Idx": 6449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904353.255, "dur": 0.081,
    "args": {
      "Python parent id": 3549, "Python id": 3550, "finished": true, "Ev Idx": 6450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904353.823, "dur": 10.014,
    "args": {
      "Python parent id": 3547, "Python id": 3551, "finished": true, "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904353.961, "dur": 0.063,
    "args": {
      "Python parent id": 3551, "Python id": 3552, "finished": true, "Ev Idx": 6452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904354.188, "dur": 0.063,
    "args": {
      "Python parent id": 3551, "Python id": 3553, "finished": true, "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904354.390, "dur": 0.057,
    "args": {
      "Python parent id": 3551, "Python id": 3554, "finished": true, "Ev Idx": 6454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904354.939, "dur": 0.077,
    "args": {
      "Python parent id": 3551, "Python id": 3555, "finished": true, "Ev Idx": 6455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904355.290, "dur": 0.070,
    "args": {
      "Python parent id": 3551, "Python id": 3556, "finished": true, "Ev Idx": 6456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904355.743, "dur": 8.005,
    "args": {
      "Python parent id": 3551, "Python id": 3557, "finished": true, "Ev Idx": 6457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904355.837, "dur": 0.061,
    "args": {
      "Python parent id": 3557, "Python id": 3558, "finished": true, "Ev Idx": 6458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904356.082, "dur": 0.058,
    "args": {
      "Python parent id": 3557, "Python id": 3559, "finished": true, "Ev Idx": 6459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904356.338, "dur": 0.072,
    "args": {
      "Python parent id": 3557, "Python id": 3560, "finished": true, "Ev Idx": 6460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904356.823, "dur": 0.034,
    "args": {
      "Python parent id": 3557, "Python id": 3561, "finished": true, "Ev Idx": 6461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904357.232, "dur": 6.456,
    "args": {
      "Python parent id": 3557, "Python id": 3562, "finished": true, "Ev Idx": 6462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904357.585, "dur": 6.022,
    "args": {
      "Python parent id": 3562, "Python id": 3563, "finished": true, "Ev Idx": 6463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904357.749, "dur": 4.028,
    "args": {
      "Python parent id": 3563, "Python id": 3564, "finished": true, "Ev Idx": 6464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904357.948, "dur": 3.764,
    "args": {
      "Python parent id": 3564, "Python id": 3565, "finished": true, "Ev Idx": 6465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904358.237, "dur": 0.279,
    "args": {
      "Python parent id": 3565, "Python id": 3566, "finished": true, "Ev Idx": 6466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904358.391, "dur": 0.070,
    "args": {
      "Python parent id": 3566, "Python id": 3567, "finished": true, "Ev Idx": 6467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904358.756, "dur": 2.499,
    "args": {
      "Python parent id": 3565, "Python id": 3568, "finished": true, "Ev Idx": 6468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.128, "dur": 2.019,
    "args": {
      "Python parent id": 3568, "Python id": 3569, "finished": true, "Ev Idx": 6469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.341, "dur": 1.732,
    "args": {
      "Python parent id": 3569, "Python id": 3570, "finished": true, "Ev Idx": 6470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.494, "dur": 1.338,
    "args": {
      "Python parent id": 3570, "Python id": 3571, "finished": true, "Ev Idx": 6471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.693, "dur": 0.579,
    "args": {
      "Python parent id": 3571, "Python id": 3572, "finished": true, "Ev Idx": 6472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.783, "dur": 0.058,
    "args": {
      "Python parent id": 3572, "Python id": 3573, "finished": true, "Ev Idx": 6473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904359.996, "dur": 0.208,
    "args": {
      "Python parent id": 3572, "Python id": 3574, "finished": true, "Ev Idx": 6474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904361.517, "dur": 0.142,
    "args": {
      "Python parent id": 3565, "Python id": 3575, "finished": true, "Ev Idx": 6475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904361.966, "dur": 0.071,
    "args": {
      "Python parent id": 3563, "Python id": 3576, "finished": true, "Ev Idx": 6476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904362.348, "dur": 1.176,
    "args": {
      "Python parent id": 3563, "Python id": 3577, "finished": true, "Ev Idx": 6477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904362.482, "dur": 1.001,
    "args": {
      "Python parent id": 3577, "Python id": 3578, "finished": true, "Ev Idx": 6478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904362.620, "dur": 0.403,
    "args": {
      "Python parent id": 3578, "Python id": 3579, "finished": true, "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904362.794, "dur": 0.170,
    "args": {
      "Python parent id": 3579, "Python id": 3580, "finished": true, "Ev Idx": 6480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904362.866, "dur": 0.060,
    "args": {
      "Python parent id": 3580, "Python id": 3581, "finished": true, "Ev Idx": 6481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904363.157, "dur": 0.271,
    "args": {
      "Python parent id": 3578, "Python id": 3582, "finished": true, "Ev Idx": 6482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904363.959, "dur": 0.243,
    "args": {
      "Python parent id": 3547, "Python id": 3583, "finished": true, "Ev Idx": 6483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904364.897, "dur": 1.164,
    "args": {
      "Python parent id": 3547, "Python id": 3584, "finished": true, "Ev Idx": 6484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904365.360, "dur": 0.648,
    "args": {
      "Python parent id": 3584, "Python id": 3585, "finished": true, "Ev Idx": 6485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904366.568, "dur": 4.561,
    "args": {
      "Python parent id": 3546, "Python id": 3586, "finished": true, "Ev Idx": 6486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904371.746, "dur": 6.868,
    "args": {
      "Python parent id": 12, "Python id": 3587, "finished": true, "Ev Idx": 6487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904379.097, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 3588, "finished": true, "Ev Idx": 6488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904379.820, "dur": 0.179,
    "args": {
      "Python parent id": 12, "Python id": 3589, "finished": true, "Ev Idx": 6489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904380.476, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 3590, "finished": true, "Ev Idx": 6490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904380.954, "dur": 9.584,
    "args": {
      "Python parent id": 12, "Python id": 3591, "finished": true, "Ev Idx": 6491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904391.722, "dur": 16.860,
    "args": {
      "Python parent id": 12, "Python id": 3592, "finished": true, "Ev Idx": 6492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904392.034, "dur": 13.423,
    "args": {
      "Python parent id": 3592, "Python id": 3593, "finished": true, "Ev Idx": 6493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904392.181, "dur": 0.478,
    "args": {
      "Python parent id": 3593, "Python id": 3594, "finished": true, "Ev Idx": 6494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904392.355, "dur": 0.227,
    "args": {
      "Python parent id": 3594, "Python id": 3595, "finished": true, "Ev Idx": 6495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904392.451, "dur": 0.097,
    "args": {
      "Python parent id": 3595, "Python id": 3596, "finished": true, "Ev Idx": 6496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904392.933, "dur": 10.481,
    "args": {
      "Python parent id": 3593, "Python id": 3597, "finished": true, "Ev Idx": 6497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904393.022, "dur": 0.107,
    "args": {
      "Python parent id": 3597, "Python id": 3598, "finished": true, "Ev Idx": 6498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904393.241, "dur": 0.063,
    "args": {
      "Python parent id": 3597, "Python id": 3599, "finished": true, "Ev Idx": 6499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904393.500, "dur": 0.060,
    "args": {
      "Python parent id": 3597, "Python id": 3600, "finished": true, "Ev Idx": 6500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904394.042, "dur": 0.036,
    "args": {
      "Python parent id": 3597, "Python id": 3601, "finished": true, "Ev Idx": 6501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904394.367, "dur": 1.266,
    "args": {
      "Python parent id": 3597, "Python id": 3602, "finished": true, "Ev Idx": 6502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904396.005, "dur": 7.332,
    "args": {
      "Python parent id": 3597, "Python id": 3603, "finished": true, "Ev Idx": 6503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904396.110, "dur": 0.062,
    "args": {
      "Python parent id": 3603, "Python id": 3604, "finished": true, "Ev Idx": 6504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904396.323, "dur": 0.058,
    "args": {
      "Python parent id": 3603, "Python id": 3605, "finished": true, "Ev Idx": 6505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904396.584, "dur": 0.070,
    "args": {
      "Python parent id": 3603, "Python id": 3606, "finished": true, "Ev Idx": 6506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904396.967, "dur": 0.043,
    "args": {
      "Python parent id": 3603, "Python id": 3607, "finished": true, "Ev Idx": 6507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904397.382, "dur": 5.856,
    "args": {
      "Python parent id": 3603, "Python id": 3608, "finished": true, "Ev Idx": 6508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904397.612, "dur": 5.562,
    "args": {
      "Python parent id": 3608, "Python id": 3609, "finished": true, "Ev Idx": 6509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904397.754, "dur": 3.635,
    "args": {
      "Python parent id": 3609, "Python id": 3610, "finished": true, "Ev Idx": 6510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904398.005, "dur": 3.294,
    "args": {
      "Python parent id": 3610, "Python id": 3611, "finished": true, "Ev Idx": 6511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904398.179, "dur": 0.308,
    "args": {
      "Python parent id": 3611, "Python id": 3612, "finished": true, "Ev Idx": 6512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904398.322, "dur": 0.065,
    "args": {
      "Python parent id": 3612, "Python id": 3613, "finished": true, "Ev Idx": 6513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904398.719, "dur": 2.221,
    "args": {
      "Python parent id": 3611, "Python id": 3614, "finished": true, "Ev Idx": 6514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904398.977, "dur": 1.906,
    "args": {
      "Python parent id": 3614, "Python id": 3615, "finished": true, "Ev Idx": 6515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904399.198, "dur": 1.623,
    "args": {
      "Python parent id": 3615, "Python id": 3616, "finished": true, "Ev Idx": 6516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904399.353, "dur": 1.190,
    "args": {
      "Python parent id": 3616, "Python id": 3617, "finished": true, "Ev Idx": 6517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904399.572, "dur": 0.442,
    "args": {
      "Python parent id": 3617, "Python id": 3618, "finished": true, "Ev Idx": 6518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904399.678, "dur": 0.047,
    "args": {
      "Python parent id": 3618, "Python id": 3619, "finished": true, "Ev Idx": 6519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904399.857, "dur": 0.117,
    "args": {
      "Python parent id": 3618, "Python id": 3620, "finished": true, "Ev Idx": 6520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904401.165, "dur": 0.071,
    "args": {
      "Python parent id": 3611, "Python id": 3621, "finished": true, "Ev Idx": 6521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904401.517, "dur": 0.073,
    "args": {
      "Python parent id": 3609, "Python id": 3622, "finished": true, "Ev Idx": 6522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904401.850, "dur": 1.276,
    "args": {
      "Python parent id": 3609, "Python id": 3623, "finished": true, "Ev Idx": 6523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904402.072, "dur": 1.006,
    "args": {
      "Python parent id": 3623, "Python id": 3624, "finished": true, "Ev Idx": 6524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904402.249, "dur": 0.388,
    "args": {
      "Python parent id": 3624, "Python id": 3625, "finished": true, "Ev Idx": 6525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904402.386, "dur": 0.187,
    "args": {
      "Python parent id": 3625, "Python id": 3626, "finished": true, "Ev Idx": 6526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904402.482, "dur": 0.061,
    "args": {
      "Python parent id": 3626, "Python id": 3627, "finished": true, "Ev Idx": 6527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904402.762, "dur": 0.264,
    "args": {
      "Python parent id": 3624, "Python id": 3628, "finished": true, "Ev Idx": 6528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904403.550, "dur": 0.205,
    "args": {
      "Python parent id": 3593, "Python id": 3629, "finished": true, "Ev Idx": 6529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904404.355, "dur": 0.951,
    "args": {
      "Python parent id": 3593, "Python id": 3630, "finished": true, "Ev Idx": 6530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904404.754, "dur": 0.496,
    "args": {
      "Python parent id": 3630, "Python id": 3631, "finished": true, "Ev Idx": 6531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904405.720, "dur": 2.815,
    "args": {
      "Python parent id": 3592, "Python id": 3632, "finished": true, "Ev Idx": 6532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904409.545, "dur": 86.123,
    "args": {
      "Python parent id": 12, "Python id": 3633, "finished": true, "Ev Idx": 6533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904409.921, "dur": 0.386,
    "args": {
      "Python parent id": 3633, "Python id": 3634, "finished": true, "Ev Idx": 6534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904410.171, "dur": 0.087,
    "args": {
      "Python parent id": 3634, "Python id": 3635, "finished": true, "Ev Idx": 6535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904410.592, "dur": 0.357,
    "args": {
      "Python parent id": 3633, "Python id": 3636, "finished": true, "Ev Idx": 6536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904411.199, "dur": 0.302,
    "args": {
      "Python parent id": 3633, "Python id": 3637, "finished": true, "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904411.754, "dur": 1.671,
    "args": {
      "Python parent id": 3633, "Python id": 3638, "finished": true, "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904414.003, "dur": 81.571,
    "args": {
      "Python parent id": 3633, "Python id": 3639, "finished": true, "Ev Idx": 6539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904414.887, "dur": 80.452,
    "args": {
      "Python parent id": 3639, "Python id": 3640, "finished": true, "Ev Idx": 6540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904415.181, "dur": 80.074,
    "args": {
      "Python parent id": 3640, "Python id": 3641, "finished": true, "Ev Idx": 6541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904420.663, "dur": 73.243,
    "args": {
      "Python parent id": 3641, "Python id": 3642, "finished": true, "Ev Idx": 6542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904421.019, "dur": 72.807,
    "args": {
      "Python parent id": 3642, "Python id": 3643, "finished": true, "Ev Idx": 6543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904421.585, "dur": 72.098,
    "args": {
      "Python parent id": 3643, "Python id": 3644, "finished": true, "Ev Idx": 6544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904422.067, "dur": 0.372,
    "args": {
      "Python parent id": 3644, "Python id": 3645, "finished": true, "Ev Idx": 6545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904422.319, "dur": 0.059,
    "args": {
      "Python parent id": 3645, "Python id": 3646, "finished": true, "Ev Idx": 6546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904422.740, "dur": 0.329,
    "args": {
      "Python parent id": 3644, "Python id": 3647, "finished": true, "Ev Idx": 6547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904422.949, "dur": 0.046,
    "args": {
      "Python parent id": 3647, "Python id": 3648, "finished": true, "Ev Idx": 6548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904423.432, "dur": 70.146,
    "args": {
      "Python parent id": 3644, "Python id": 3649, "finished": true, "Ev Idx": 6549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904424.056, "dur": 68.272,
    "args": {
      "Python parent id": 3649, "Python id": 3650, "finished": true, "Ev Idx": 6550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904431.822, "dur": 17.095,
    "args": {
      "Python parent id": 3650, "Python id": 3651, "finished": true, "Ev Idx": 6551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904432.039, "dur": 0.105,
    "args": {
      "Python parent id": 3651, "Python id": 3652, "finished": true, "Ev Idx": 6552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904432.499, "dur": 16.312,
    "args": {
      "Python parent id": 3651, "Python id": 3653, "finished": true, "Ev Idx": 6553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904432.690, "dur": 7.298,
    "args": {
      "Python parent id": 3653, "Python id": 3654, "finished": true, "Ev Idx": 6554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904456.368, "dur": 6.535,
    "args": {
      "Python parent id": 3650, "Python id": 3655, "finished": true, "Ev Idx": 6555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904470.747, "dur": 7.869,
    "args": {
      "Python parent id": 3650, "Python id": 3656, "finished": true, "Ev Idx": 6556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904496.879, "dur": 16.659,
    "args": {
      "Python parent id": 12, "Python id": 3657, "finished": true, "Ev Idx": 6557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904497.235, "dur": 12.468,
    "args": {
      "Python parent id": 3657, "Python id": 3658, "finished": true, "Ev Idx": 6558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904497.423, "dur": 0.457,
    "args": {
      "Python parent id": 3658, "Python id": 3659, "finished": true, "Ev Idx": 6559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904497.595, "dur": 0.215,
    "args": {
      "Python parent id": 3659, "Python id": 3660, "finished": true, "Ev Idx": 6560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904497.698, "dur": 0.076,
    "args": {
      "Python parent id": 3660, "Python id": 3661, "finished": true, "Ev Idx": 6561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904498.274, "dur": 9.246,
    "args": {
      "Python parent id": 3658, "Python id": 3662, "finished": true, "Ev Idx": 6562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904498.379, "dur": 0.064,
    "args": {
      "Python parent id": 3662, "Python id": 3663, "finished": true, "Ev Idx": 6563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904498.551, "dur": 0.061,
    "args": {
      "Python parent id": 3662, "Python id": 3664, "finished": true, "Ev Idx": 6564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904498.752, "dur": 0.058,
    "args": {
      "Python parent id": 3662, "Python id": 3665, "finished": true, "Ev Idx": 6565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904499.277, "dur": 0.030,
    "args": {
      "Python parent id": 3662, "Python id": 3666, "finished": true, "Ev Idx": 6566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904499.582, "dur": 0.074,
    "args": {
      "Python parent id": 3662, "Python id": 3667, "finished": true, "Ev Idx": 6567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904500.075, "dur": 7.366,
    "args": {
      "Python parent id": 3662, "Python id": 3668, "finished": true, "Ev Idx": 6568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904500.168, "dur": 0.064,
    "args": {
      "Python parent id": 3668, "Python id": 3669, "finished": true, "Ev Idx": 6569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904500.374, "dur": 0.058,
    "args": {
      "Python parent id": 3668, "Python id": 3670, "finished": true, "Ev Idx": 6570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904500.626, "dur": 0.070,
    "args": {
      "Python parent id": 3668, "Python id": 3671, "finished": true, "Ev Idx": 6571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904501.058, "dur": 0.037,
    "args": {
      "Python parent id": 3668, "Python id": 3672, "finished": true, "Ev Idx": 6572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904501.438, "dur": 5.914,
    "args": {
      "Python parent id": 3668, "Python id": 3673, "finished": true, "Ev Idx": 6573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904501.729, "dur": 5.543,
    "args": {
      "Python parent id": 3673, "Python id": 3674, "finished": true, "Ev Idx": 6574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904501.856, "dur": 3.509,
    "args": {
      "Python parent id": 3674, "Python id": 3675, "finished": true, "Ev Idx": 6575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904502.047, "dur": 3.271,
    "args": {
      "Python parent id": 3675, "Python id": 3676, "finished": true, "Ev Idx": 6576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904502.221, "dur": 0.255,
    "args": {
      "Python parent id": 3676, "Python id": 3677, "finished": true, "Ev Idx": 6577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904502.347, "dur": 0.069,
    "args": {
      "Python parent id": 3677, "Python id": 3678, "finished": true, "Ev Idx": 6578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904502.670, "dur": 2.298,
    "args": {
      "Python parent id": 3676, "Python id": 3679, "finished": true, "Ev Idx": 6579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904502.925, "dur": 1.958,
    "args": {
      "Python parent id": 3679, "Python id": 3680, "finished": true, "Ev Idx": 6580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904503.165, "dur": 1.670,
    "args": {
      "Python parent id": 3680, "Python id": 3681, "finished": true, "Ev Idx": 6581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904503.351, "dur": 1.226,
    "args": {
      "Python parent id": 3681, "Python id": 3682, "finished": true, "Ev Idx": 6582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904503.590, "dur": 0.485,
    "args": {
      "Python parent id": 3682, "Python id": 3683, "finished": true, "Ev Idx": 6583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904503.677, "dur": 0.050,
    "args": {
      "Python parent id": 3683, "Python id": 3684, "finished": true, "Ev Idx": 6584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904503.907, "dur": 0.113,
    "args": {
      "Python parent id": 3683, "Python id": 3685, "finished": true, "Ev Idx": 6585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904505.152, "dur": 0.120,
    "args": {
      "Python parent id": 3676, "Python id": 3686, "finished": true, "Ev Idx": 6586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904505.530, "dur": 0.112,
    "args": {
      "Python parent id": 3674, "Python id": 3687, "finished": true, "Ev Idx": 6587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904505.955, "dur": 1.226,
    "args": {
      "Python parent id": 3674, "Python id": 3688, "finished": true, "Ev Idx": 6588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904506.115, "dur": 1.021,
    "args": {
      "Python parent id": 3688, "Python id": 3689, "finished": true, "Ev Idx": 6589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904506.281, "dur": 0.382,
    "args": {
      "Python parent id": 3689, "Python id": 3690, "finished": true, "Ev Idx": 6590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904506.419, "dur": 0.185,
    "args": {
      "Python parent id": 3690, "Python id": 3691, "finished": true, "Ev Idx": 6591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904506.513, "dur": 0.055,
    "args": {
      "Python parent id": 3691, "Python id": 3692, "finished": true, "Ev Idx": 6592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904506.802, "dur": 0.298,
    "args": {
      "Python parent id": 3689, "Python id": 3693, "finished": true, "Ev Idx": 6593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904507.641, "dur": 0.230,
    "args": {
      "Python parent id": 3658, "Python id": 3694, "finished": true, "Ev Idx": 6594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904508.493, "dur": 1.064,
    "args": {
      "Python parent id": 3658, "Python id": 3695, "finished": true, "Ev Idx": 6595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904508.952, "dur": 0.552,
    "args": {
      "Python parent id": 3695, "Python id": 3696, "finished": true, "Ev Idx": 6596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904510.045, "dur": 3.409,
    "args": {
      "Python parent id": 3657, "Python id": 3697, "finished": true, "Ev Idx": 6597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904514.167, "dur": 6.897,
    "args": {
      "Python parent id": 12, "Python id": 3698, "finished": true, "Ev Idx": 6598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904521.539, "dur": 0.084,
    "args": {
      "Python parent id": 12, "Python id": 3699, "finished": true, "Ev Idx": 6599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904522.248, "dur": 0.187,
    "args": {
      "Python parent id": 12, "Python id": 3700, "finished": true, "Ev Idx": 6600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904522.922, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 3701, "finished": true, "Ev Idx": 6601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904523.432, "dur": 9.572,
    "args": {
      "Python parent id": 12, "Python id": 3702, "finished": true, "Ev Idx": 6602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904534.280, "dur": 14.897,
    "args": {
      "Python parent id": 12, "Python id": 3703, "finished": true, "Ev Idx": 6603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904534.551, "dur": 11.440,
    "args": {
      "Python parent id": 3703, "Python id": 3704, "finished": true, "Ev Idx": 6604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904534.702, "dur": 0.502,
    "args": {
      "Python parent id": 3704, "Python id": 3705, "finished": true, "Ev Idx": 6605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904534.872, "dur": 0.232,
    "args": {
      "Python parent id": 3705, "Python id": 3706, "finished": true, "Ev Idx": 6606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904534.974, "dur": 0.095,
    "args": {
      "Python parent id": 3706, "Python id": 3707, "finished": true, "Ev Idx": 6607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904535.487, "dur": 8.673,
    "args": {
      "Python parent id": 3704, "Python id": 3708, "finished": true, "Ev Idx": 6608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904535.573, "dur": 0.099,
    "args": {
      "Python parent id": 3708, "Python id": 3709, "finished": true, "Ev Idx": 6609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904535.794, "dur": 0.078,
    "args": {
      "Python parent id": 3708, "Python id": 3710, "finished": true, "Ev Idx": 6610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904536.007, "dur": 0.059,
    "args": {
      "Python parent id": 3708, "Python id": 3711, "finished": true, "Ev Idx": 6611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904536.498, "dur": 0.038,
    "args": {
      "Python parent id": 3708, "Python id": 3712, "finished": true, "Ev Idx": 6612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904536.781, "dur": 0.079,
    "args": {
      "Python parent id": 3708, "Python id": 3713, "finished": true, "Ev Idx": 6613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904537.224, "dur": 6.873,
    "args": {
      "Python parent id": 3708, "Python id": 3714, "finished": true, "Ev Idx": 6614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904537.326, "dur": 0.065,
    "args": {
      "Python parent id": 3714, "Python id": 3715, "finished": true, "Ev Idx": 6615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904537.536, "dur": 0.067,
    "args": {
      "Python parent id": 3714, "Python id": 3716, "finished": true, "Ev Idx": 6616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904537.813, "dur": 0.071,
    "args": {
      "Python parent id": 3714, "Python id": 3717, "finished": true, "Ev Idx": 6617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904538.256, "dur": 0.036,
    "args": {
      "Python parent id": 3714, "Python id": 3718, "finished": true, "Ev Idx": 6618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904538.623, "dur": 5.401,
    "args": {
      "Python parent id": 3714, "Python id": 3719, "finished": true, "Ev Idx": 6619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904538.884, "dur": 5.083,
    "args": {
      "Python parent id": 3719, "Python id": 3720, "finished": true, "Ev Idx": 6620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.062, "dur": 3.285,
    "args": {
      "Python parent id": 3720, "Python id": 3721, "finished": true, "Ev Idx": 6621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.240, "dur": 3.055,
    "args": {
      "Python parent id": 3721, "Python id": 3722, "finished": true, "Ev Idx": 6622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.401, "dur": 0.249,
    "args": {
      "Python parent id": 3722, "Python id": 3723, "finished": true, "Ev Idx": 6623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.514, "dur": 0.074,
    "args": {
      "Python parent id": 3723, "Python id": 3724, "finished": true, "Ev Idx": 6624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.802, "dur": 2.116,
    "args": {
      "Python parent id": 3722, "Python id": 3725, "finished": true, "Ev Idx": 6625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904539.981, "dur": 1.752,
    "args": {
      "Python parent id": 3725, "Python id": 3726, "finished": true, "Ev Idx": 6626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904540.171, "dur": 1.512,
    "args": {
      "Python parent id": 3726, "Python id": 3727, "finished": true, "Ev Idx": 6627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904540.343, "dur": 1.123,
    "args": {
      "Python parent id": 3727, "Python id": 3728, "finished": true, "Ev Idx": 6628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904540.531, "dur": 0.475,
    "args": {
      "Python parent id": 3728, "Python id": 3729, "finished": true, "Ev Idx": 6629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904540.620, "dur": 0.047,
    "args": {
      "Python parent id": 3729, "Python id": 3730, "finished": true, "Ev Idx": 6630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904540.824, "dur": 0.143,
    "args": {
      "Python parent id": 3729, "Python id": 3731, "finished": true, "Ev Idx": 6631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904542.202, "dur": 0.059,
    "args": {
      "Python parent id": 3722, "Python id": 3732, "finished": true, "Ev Idx": 6632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904542.492, "dur": 0.073,
    "args": {
      "Python parent id": 3720, "Python id": 3733, "finished": true, "Ev Idx": 6633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904542.817, "dur": 1.105,
    "args": {
      "Python parent id": 3720, "Python id": 3734, "finished": true, "Ev Idx": 6634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904542.955, "dur": 0.922,
    "args": {
      "Python parent id": 3734, "Python id": 3735, "finished": true, "Ev Idx": 6635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904543.103, "dur": 0.369,
    "args": {
      "Python parent id": 3735, "Python id": 3736, "finished": true, "Ev Idx": 6636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904543.240, "dur": 0.172,
    "args": {
      "Python parent id": 3736, "Python id": 3737, "finished": true, "Ev Idx": 6637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904543.328, "dur": 0.052,
    "args": {
      "Python parent id": 3737, "Python id": 3738, "finished": true, "Ev Idx": 6638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904543.594, "dur": 0.253,
    "args": {
      "Python parent id": 3735, "Python id": 3739, "finished": true, "Ev Idx": 6639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904544.282, "dur": 0.172,
    "args": {
      "Python parent id": 3704, "Python id": 3740, "finished": true, "Ev Idx": 6640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904544.989, "dur": 0.866,
    "args": {
      "Python parent id": 3704, "Python id": 3741, "finished": true, "Ev Idx": 6641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904545.365, "dur": 0.442,
    "args": {
      "Python parent id": 3741, "Python id": 3742, "finished": true, "Ev Idx": 6642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904546.232, "dur": 2.894,
    "args": {
      "Python parent id": 3703, "Python id": 3743, "finished": true, "Ev Idx": 6643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904550.089, "dur": 89.441,
    "args": {
      "Python parent id": 12, "Python id": 3744, "finished": true, "Ev Idx": 6644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904550.449, "dur": 0.407,
    "args": {
      "Python parent id": 3744, "Python id": 3745, "finished": true, "Ev Idx": 6645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904550.717, "dur": 0.086,
    "args": {
      "Python parent id": 3745, "Python id": 3746, "finished": true, "Ev Idx": 6646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904551.189, "dur": 0.363,
    "args": {
      "Python parent id": 3744, "Python id": 3747, "finished": true, "Ev Idx": 6647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904551.823, "dur": 0.263,
    "args": {
      "Python parent id": 3744, "Python id": 3748, "finished": true, "Ev Idx": 6648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904552.324, "dur": 1.510,
    "args": {
      "Python parent id": 3744, "Python id": 3749, "finished": true, "Ev Idx": 6649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904554.389, "dur": 84.995,
    "args": {
      "Python parent id": 3744, "Python id": 3750, "finished": true, "Ev Idx": 6650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904555.330, "dur": 83.853,
    "args": {
      "Python parent id": 3750, "Python id": 3751, "finished": true, "Ev Idx": 6651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904555.565, "dur": 83.555,
    "args": {
      "Python parent id": 3751, "Python id": 3752, "finished": true, "Ev Idx": 6652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904560.906, "dur": 76.927,
    "args": {
      "Python parent id": 3752, "Python id": 3753, "finished": true, "Ev Idx": 6653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904561.221, "dur": 76.468,
    "args": {
      "Python parent id": 3753, "Python id": 3754, "finished": true, "Ev Idx": 6654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904561.784, "dur": 75.824,
    "args": {
      "Python parent id": 3754, "Python id": 3755, "finished": true, "Ev Idx": 6655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904562.249, "dur": 0.430,
    "args": {
      "Python parent id": 3755, "Python id": 3756, "finished": true, "Ev Idx": 6656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904562.534, "dur": 0.057,
    "args": {
      "Python parent id": 3756, "Python id": 3757, "finished": true, "Ev Idx": 6657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904563.019, "dur": 0.347,
    "args": {
      "Python parent id": 3755, "Python id": 3758, "finished": true, "Ev Idx": 6658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904563.261, "dur": 0.051,
    "args": {
      "Python parent id": 3758, "Python id": 3759, "finished": true, "Ev Idx": 6659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904563.754, "dur": 73.790,
    "args": {
      "Python parent id": 3755, "Python id": 3760, "finished": true, "Ev Idx": 6660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904564.430, "dur": 71.868,
    "args": {
      "Python parent id": 3760, "Python id": 3761, "finished": true, "Ev Idx": 6661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904571.788, "dur": 18.318,
    "args": {
      "Python parent id": 3761, "Python id": 3762, "finished": true, "Ev Idx": 6662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904571.935, "dur": 0.108,
    "args": {
      "Python parent id": 3762, "Python id": 3763, "finished": true, "Ev Idx": 6663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904572.333, "dur": 17.676,
    "args": {
      "Python parent id": 3762, "Python id": 3764, "finished": true, "Ev Idx": 6664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904572.491, "dur": 8.052,
    "args": {
      "Python parent id": 3764, "Python id": 3765, "finished": true, "Ev Idx": 6665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904597.534, "dur": 6.588,
    "args": {
      "Python parent id": 3761, "Python id": 3766, "finished": true, "Ev Idx": 6666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904613.144, "dur": 8.168,
    "args": {
      "Python parent id": 3761, "Python id": 3767, "finished": true, "Ev Idx": 6667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904640.803, "dur": 17.888,
    "args": {
      "Python parent id": 12, "Python id": 3768, "finished": true, "Ev Idx": 6668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904641.221, "dur": 13.334,
    "args": {
      "Python parent id": 3768, "Python id": 3769, "finished": true, "Ev Idx": 6669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904641.407, "dur": 0.494,
    "args": {
      "Python parent id": 3769, "Python id": 3770, "finished": true, "Ev Idx": 6670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904641.597, "dur": 0.219,
    "args": {
      "Python parent id": 3770, "Python id": 3771, "finished": true, "Ev Idx": 6671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904641.696, "dur": 0.085,
    "args": {
      "Python parent id": 3771, "Python id": 3772, "finished": true, "Ev Idx": 6672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904642.256, "dur": 10.049,
    "args": {
      "Python parent id": 3769, "Python id": 3773, "finished": true, "Ev Idx": 6673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904642.399, "dur": 0.065,
    "args": {
      "Python parent id": 3773, "Python id": 3774, "finished": true, "Ev Idx": 6674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904642.578, "dur": 0.066,
    "args": {
      "Python parent id": 3773, "Python id": 3775, "finished": true, "Ev Idx": 6675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904642.831, "dur": 0.062,
    "args": {
      "Python parent id": 3773, "Python id": 3776, "finished": true, "Ev Idx": 6676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904643.386, "dur": 0.059,
    "args": {
      "Python parent id": 3773, "Python id": 3777, "finished": true, "Ev Idx": 6677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904643.793, "dur": 0.092,
    "args": {
      "Python parent id": 3773, "Python id": 3778, "finished": true, "Ev Idx": 6678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904644.308, "dur": 7.916,
    "args": {
      "Python parent id": 3773, "Python id": 3779, "finished": true, "Ev Idx": 6679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904644.424, "dur": 0.062,
    "args": {
      "Python parent id": 3779, "Python id": 3780, "finished": true, "Ev Idx": 6680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904644.680, "dur": 0.061,
    "args": {
      "Python parent id": 3779, "Python id": 3781, "finished": true, "Ev Idx": 6681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904644.938, "dur": 0.071,
    "args": {
      "Python parent id": 3779, "Python id": 3782, "finished": true, "Ev Idx": 6682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904645.376, "dur": 0.075,
    "args": {
      "Python parent id": 3779, "Python id": 3783, "finished": true, "Ev Idx": 6683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904645.796, "dur": 6.302,
    "args": {
      "Python parent id": 3779, "Python id": 3784, "finished": true, "Ev Idx": 6684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904646.081, "dur": 5.937,
    "args": {
      "Python parent id": 3784, "Python id": 3785, "finished": true, "Ev Idx": 6685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904646.248, "dur": 3.872,
    "args": {
      "Python parent id": 3785, "Python id": 3786, "finished": true, "Ev Idx": 6686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904646.442, "dur": 3.592,
    "args": {
      "Python parent id": 3786, "Python id": 3787, "finished": true, "Ev Idx": 6687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904646.607, "dur": 0.347,
    "args": {
      "Python parent id": 3787, "Python id": 3788, "finished": true, "Ev Idx": 6688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904646.769, "dur": 0.100,
    "args": {
      "Python parent id": 3788, "Python id": 3789, "finished": true, "Ev Idx": 6689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904647.177, "dur": 2.429,
    "args": {
      "Python parent id": 3787, "Python id": 3790, "finished": true, "Ev Idx": 6690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904647.387, "dur": 2.105,
    "args": {
      "Python parent id": 3790, "Python id": 3791, "finished": true, "Ev Idx": 6691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904647.616, "dur": 1.785,
    "args": {
      "Python parent id": 3791, "Python id": 3792, "finished": true, "Ev Idx": 6692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904647.780, "dur": 1.363,
    "args": {
      "Python parent id": 3792, "Python id": 3793, "finished": true, "Ev Idx": 6693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904647.965, "dur": 0.621,
    "args": {
      "Python parent id": 3793, "Python id": 3794, "finished": true, "Ev Idx": 6694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904648.066, "dur": 0.053,
    "args": {
      "Python parent id": 3794, "Python id": 3795, "finished": true, "Ev Idx": 6695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904648.272, "dur": 0.252,
    "args": {
      "Python parent id": 3794, "Python id": 3796, "finished": true, "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904649.841, "dur": 0.156,
    "args": {
      "Python parent id": 3787, "Python id": 3797, "finished": true, "Ev Idx": 6697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904650.316, "dur": 0.121,
    "args": {
      "Python parent id": 3785, "Python id": 3798, "finished": true, "Ev Idx": 6698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904650.708, "dur": 1.241,
    "args": {
      "Python parent id": 3785, "Python id": 3799, "finished": true, "Ev Idx": 6699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904650.878, "dur": 1.031,
    "args": {
      "Python parent id": 3799, "Python id": 3800, "finished": true, "Ev Idx": 6700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904651.037, "dur": 0.423,
    "args": {
      "Python parent id": 3800, "Python id": 3801, "finished": true, "Ev Idx": 6701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904651.194, "dur": 0.208,
    "args": {
      "Python parent id": 3801, "Python id": 3802, "finished": true, "Ev Idx": 6702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904651.311, "dur": 0.051,
    "args": {
      "Python parent id": 3802, "Python id": 3803, "finished": true, "Ev Idx": 6703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904651.593, "dur": 0.259,
    "args": {
      "Python parent id": 3800, "Python id": 3804, "finished": true, "Ev Idx": 6704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904652.464, "dur": 0.272,
    "args": {
      "Python parent id": 3769, "Python id": 3805, "finished": true, "Ev Idx": 6705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904653.315, "dur": 1.058,
    "args": {
      "Python parent id": 3769, "Python id": 3806, "finished": true, "Ev Idx": 6706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904653.818, "dur": 0.501,
    "args": {
      "Python parent id": 3806, "Python id": 3807, "finished": true, "Ev Idx": 6707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904654.894, "dur": 3.753,
    "args": {
      "Python parent id": 3768, "Python id": 3808, "finished": true, "Ev Idx": 6708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904659.240, "dur": 7.217,
    "args": {
      "Python parent id": 12, "Python id": 3809, "finished": true, "Ev Idx": 6709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904666.945, "dur": 0.095,
    "args": {
      "Python parent id": 12, "Python id": 3810, "finished": true, "Ev Idx": 6710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904667.662, "dur": 0.190,
    "args": {
      "Python parent id": 12, "Python id": 3811, "finished": true, "Ev Idx": 6711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904668.337, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 3812, "finished": true, "Ev Idx": 6712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904668.813, "dur": 10.253,
    "args": {
      "Python parent id": 12, "Python id": 3813, "finished": true, "Ev Idx": 6713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904680.269, "dur": 15.353,
    "args": {
      "Python parent id": 12, "Python id": 3814, "finished": true, "Ev Idx": 6714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904680.621, "dur": 11.704,
    "args": {
      "Python parent id": 3814, "Python id": 3815, "finished": true, "Ev Idx": 6715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904680.824, "dur": 0.539,
    "args": {
      "Python parent id": 3815, "Python id": 3816, "finished": true, "Ev Idx": 6716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904680.972, "dur": 0.312,
    "args": {
      "Python parent id": 3816, "Python id": 3817, "finished": true, "Ev Idx": 6717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904681.112, "dur": 0.136,
    "args": {
      "Python parent id": 3817, "Python id": 3818, "finished": true, "Ev Idx": 6718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904681.636, "dur": 8.769,
    "args": {
      "Python parent id": 3815, "Python id": 3819, "finished": true, "Ev Idx": 6719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904681.743, "dur": 0.089,
    "args": {
      "Python parent id": 3819, "Python id": 3820, "finished": true, "Ev Idx": 6720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904681.952, "dur": 0.077,
    "args": {
      "Python parent id": 3819, "Python id": 3821, "finished": true, "Ev Idx": 6721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904682.201, "dur": 0.067,
    "args": {
      "Python parent id": 3819, "Python id": 3822, "finished": true, "Ev Idx": 6722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904682.714, "dur": 0.046,
    "args": {
      "Python parent id": 3819, "Python id": 3823, "finished": true, "Ev Idx": 6723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904683.062, "dur": 0.079,
    "args": {
      "Python parent id": 3819, "Python id": 3824, "finished": true, "Ev Idx": 6724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904683.566, "dur": 6.745,
    "args": {
      "Python parent id": 3819, "Python id": 3825, "finished": true, "Ev Idx": 6725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904683.670, "dur": 0.065,
    "args": {
      "Python parent id": 3825, "Python id": 3826, "finished": true, "Ev Idx": 6726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904683.922, "dur": 0.066,
    "args": {
      "Python parent id": 3825, "Python id": 3827, "finished": true, "Ev Idx": 6727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904684.160, "dur": 0.069,
    "args": {
      "Python parent id": 3825, "Python id": 3828, "finished": true, "Ev Idx": 6728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904684.560, "dur": 0.045,
    "args": {
      "Python parent id": 3825, "Python id": 3829, "finished": true, "Ev Idx": 6729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904684.925, "dur": 5.307,
    "args": {
      "Python parent id": 3825, "Python id": 3830, "finished": true, "Ev Idx": 6730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904685.199, "dur": 4.966,
    "args": {
      "Python parent id": 3830, "Python id": 3831, "finished": true, "Ev Idx": 6731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904685.338, "dur": 3.151,
    "args": {
      "Python parent id": 3831, "Python id": 3832, "finished": true, "Ev Idx": 6732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904685.520, "dur": 2.918,
    "args": {
      "Python parent id": 3832, "Python id": 3833, "finished": true, "Ev Idx": 6733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904685.659, "dur": 0.301,
    "args": {
      "Python parent id": 3833, "Python id": 3834, "finished": true, "Ev Idx": 6734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904685.817, "dur": 0.068,
    "args": {
      "Python parent id": 3834, "Python id": 3835, "finished": true, "Ev Idx": 6735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.134, "dur": 2.023,
    "args": {
      "Python parent id": 3833, "Python id": 3836, "finished": true, "Ev Idx": 6736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.348, "dur": 1.746,
    "args": {
      "Python parent id": 3836, "Python id": 3837, "finished": true, "Ev Idx": 6737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.559, "dur": 1.485,
    "args": {
      "Python parent id": 3837, "Python id": 3838, "finished": true, "Ev Idx": 6738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.750, "dur": 1.097,
    "args": {
      "Python parent id": 3838, "Python id": 3839, "finished": true, "Ev Idx": 6739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.899, "dur": 0.479,
    "args": {
      "Python parent id": 3839, "Python id": 3840, "finished": true, "Ev Idx": 6740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904686.980, "dur": 0.072,
    "args": {
      "Python parent id": 3840, "Python id": 3841, "finished": true, "Ev Idx": 6741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904687.184, "dur": 0.152,
    "args": {
      "Python parent id": 3840, "Python id": 3842, "finished": true, "Ev Idx": 6742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904688.343, "dur": 0.063,
    "args": {
      "Python parent id": 3833, "Python id": 3843, "finished": true, "Ev Idx": 6743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904688.657, "dur": 0.076,
    "args": {
      "Python parent id": 3831, "Python id": 3844, "finished": true, "Ev Idx": 6744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.001, "dur": 1.120,
    "args": {
      "Python parent id": 3831, "Python id": 3845, "finished": true, "Ev Idx": 6745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.157, "dur": 0.920,
    "args": {
      "Python parent id": 3845, "Python id": 3846, "finished": true, "Ev Idx": 6746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.292, "dur": 0.393,
    "args": {
      "Python parent id": 3846, "Python id": 3847, "finished": true, "Ev Idx": 6747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.445, "dur": 0.180,
    "args": {
      "Python parent id": 3847, "Python id": 3848, "finished": true, "Ev Idx": 6748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.536, "dur": 0.058,
    "args": {
      "Python parent id": 3848, "Python id": 3849, "finished": true, "Ev Idx": 6749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904689.815, "dur": 0.231,
    "args": {
      "Python parent id": 3846, "Python id": 3850, "finished": true, "Ev Idx": 6750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904690.536, "dur": 0.195,
    "args": {
      "Python parent id": 3815, "Python id": 3851, "finished": true, "Ev Idx": 6751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904691.322, "dur": 0.876,
    "args": {
      "Python parent id": 3815, "Python id": 3852, "finished": true, "Ev Idx": 6752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904691.682, "dur": 0.462,
    "args": {
      "Python parent id": 3852, "Python id": 3853, "finished": true, "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904692.642, "dur": 2.933,
    "args": {
      "Python parent id": 3814, "Python id": 3854, "finished": true, "Ev Idx": 6754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904696.594, "dur": 91.258,
    "args": {
      "Python parent id": 12, "Python id": 3855, "finished": true, "Ev Idx": 6755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904696.929, "dur": 0.389,
    "args": {
      "Python parent id": 3855, "Python id": 3856, "finished": true, "Ev Idx": 6756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904697.192, "dur": 0.077,
    "args": {
      "Python parent id": 3856, "Python id": 3857, "finished": true, "Ev Idx": 6757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904697.631, "dur": 0.352,
    "args": {
      "Python parent id": 3855, "Python id": 3858, "finished": true, "Ev Idx": 6758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904698.234, "dur": 0.262,
    "args": {
      "Python parent id": 3855, "Python id": 3859, "finished": true, "Ev Idx": 6759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904698.723, "dur": 1.712,
    "args": {
      "Python parent id": 3855, "Python id": 3860, "finished": true, "Ev Idx": 6760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904701.033, "dur": 86.674,
    "args": {
      "Python parent id": 3855, "Python id": 3861, "finished": true, "Ev Idx": 6761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904701.874, "dur": 85.622,
    "args": {
      "Python parent id": 3861, "Python id": 3862, "finished": true, "Ev Idx": 6762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904702.165, "dur": 85.271,
    "args": {
      "Python parent id": 3862, "Python id": 3863, "finished": true, "Ev Idx": 6763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904708.102, "dur": 77.962,
    "args": {
      "Python parent id": 3863, "Python id": 3864, "finished": true, "Ev Idx": 6764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904708.413, "dur": 77.530,
    "args": {
      "Python parent id": 3864, "Python id": 3865, "finished": true, "Ev Idx": 6765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904708.974, "dur": 76.828,
    "args": {
      "Python parent id": 3865, "Python id": 3866, "finished": true, "Ev Idx": 6766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904709.420, "dur": 0.426,
    "args": {
      "Python parent id": 3866, "Python id": 3867, "finished": true, "Ev Idx": 6767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904709.720, "dur": 0.066,
    "args": {
      "Python parent id": 3867, "Python id": 3868, "finished": true, "Ev Idx": 6768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904710.177, "dur": 0.410,
    "args": {
      "Python parent id": 3866, "Python id": 3869, "finished": true, "Ev Idx": 6769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904710.481, "dur": 0.056,
    "args": {
      "Python parent id": 3869, "Python id": 3870, "finished": true, "Ev Idx": 6770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904710.940, "dur": 74.741,
    "args": {
      "Python parent id": 3866, "Python id": 3871, "finished": true, "Ev Idx": 6771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904711.550, "dur": 72.835,
    "args": {
      "Python parent id": 3871, "Python id": 3872, "finished": true, "Ev Idx": 6772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904719.386, "dur": 17.753,
    "args": {
      "Python parent id": 3872, "Python id": 3873, "finished": true, "Ev Idx": 6773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904719.517, "dur": 0.143,
    "args": {
      "Python parent id": 3873, "Python id": 3874, "finished": true, "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904719.985, "dur": 17.064,
    "args": {
      "Python parent id": 3873, "Python id": 3875, "finished": true, "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904720.165, "dur": 7.607,
    "args": {
      "Python parent id": 3875, "Python id": 3876, "finished": true, "Ev Idx": 6776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904745.507, "dur": 7.098,
    "args": {
      "Python parent id": 3872, "Python id": 3877, "finished": true, "Ev Idx": 6777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904761.281, "dur": 8.303,
    "args": {
      "Python parent id": 3872, "Python id": 3878, "finished": true, "Ev Idx": 6778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904789.098, "dur": 16.572,
    "args": {
      "Python parent id": 12, "Python id": 3879, "finished": true, "Ev Idx": 6779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904789.543, "dur": 12.566,
    "args": {
      "Python parent id": 3879, "Python id": 3880, "finished": true, "Ev Idx": 6780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904789.835, "dur": 0.497,
    "args": {
      "Python parent id": 3880, "Python id": 3881, "finished": true, "Ev Idx": 6781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904790.024, "dur": 0.234,
    "args": {
      "Python parent id": 3881, "Python id": 3882, "finished": true, "Ev Idx": 6782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904790.140, "dur": 0.078,
    "args": {
      "Python parent id": 3882, "Python id": 3883, "finished": true, "Ev Idx": 6783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904790.636, "dur": 9.418,
    "args": {
      "Python parent id": 3880, "Python id": 3884, "finished": true, "Ev Idx": 6784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904790.788, "dur": 0.108,
    "args": {
      "Python parent id": 3884, "Python id": 3885, "finished": true, "Ev Idx": 6785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904791.022, "dur": 0.062,
    "args": {
      "Python parent id": 3884, "Python id": 3886, "finished": true, "Ev Idx": 6786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904791.247, "dur": 0.061,
    "args": {
      "Python parent id": 3884, "Python id": 3887, "finished": true, "Ev Idx": 6787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904791.793, "dur": 0.031,
    "args": {
      "Python parent id": 3884, "Python id": 3888, "finished": true, "Ev Idx": 6788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904792.135, "dur": 0.079,
    "args": {
      "Python parent id": 3884, "Python id": 3889, "finished": true, "Ev Idx": 6789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904792.586, "dur": 7.384,
    "args": {
      "Python parent id": 3884, "Python id": 3890, "finished": true, "Ev Idx": 6790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904792.704, "dur": 0.064,
    "args": {
      "Python parent id": 3890, "Python id": 3891, "finished": true, "Ev Idx": 6791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904792.944, "dur": 0.059,
    "args": {
      "Python parent id": 3890, "Python id": 3892, "finished": true, "Ev Idx": 6792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904793.234, "dur": 0.069,
    "args": {
      "Python parent id": 3890, "Python id": 3893, "finished": true, "Ev Idx": 6793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904793.706, "dur": 0.039,
    "args": {
      "Python parent id": 3890, "Python id": 3894, "finished": true, "Ev Idx": 6794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904794.125, "dur": 5.734,
    "args": {
      "Python parent id": 3890, "Python id": 3895, "finished": true, "Ev Idx": 6795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904794.442, "dur": 5.353,
    "args": {
      "Python parent id": 3895, "Python id": 3896, "finished": true, "Ev Idx": 6796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904794.581, "dur": 3.459,
    "args": {
      "Python parent id": 3896, "Python id": 3897, "finished": true, "Ev Idx": 6797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904794.765, "dur": 3.230,
    "args": {
      "Python parent id": 3897, "Python id": 3898, "finished": true, "Ev Idx": 6798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904794.938, "dur": 0.293,
    "args": {
      "Python parent id": 3898, "Python id": 3899, "finished": true, "Ev Idx": 6799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904795.089, "dur": 0.061,
    "args": {
      "Python parent id": 3899, "Python id": 3900, "finished": true, "Ev Idx": 6800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904795.418, "dur": 2.182,
    "args": {
      "Python parent id": 3898, "Python id": 3901, "finished": true, "Ev Idx": 6801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904795.657, "dur": 1.864,
    "args": {
      "Python parent id": 3901, "Python id": 3902, "finished": true, "Ev Idx": 6802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904795.875, "dur": 1.599,
    "args": {
      "Python parent id": 3902, "Python id": 3903, "finished": true, "Ev Idx": 6803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904796.028, "dur": 1.231,
    "args": {
      "Python parent id": 3903, "Python id": 3904, "finished": true, "Ev Idx": 6804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904796.235, "dur": 0.533,
    "args": {
      "Python parent id": 3904, "Python id": 3905, "finished": true, "Ev Idx": 6805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904796.321, "dur": 0.051,
    "args": {
      "Python parent id": 3905, "Python id": 3906, "finished": true, "Ev Idx": 6806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904796.551, "dur": 0.151,
    "args": {
      "Python parent id": 3905, "Python id": 3907, "finished": true, "Ev Idx": 6807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904797.806, "dur": 0.143,
    "args": {
      "Python parent id": 3898, "Python id": 3908, "finished": true, "Ev Idx": 6808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904798.196, "dur": 0.103,
    "args": {
      "Python parent id": 3896, "Python id": 3909, "finished": true, "Ev Idx": 6809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904798.558, "dur": 1.190,
    "args": {
      "Python parent id": 3896, "Python id": 3910, "finished": true, "Ev Idx": 6810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904798.727, "dur": 0.975,
    "args": {
      "Python parent id": 3910, "Python id": 3911, "finished": true, "Ev Idx": 6811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904798.873, "dur": 0.368,
    "args": {
      "Python parent id": 3911, "Python id": 3912, "finished": true, "Ev Idx": 6812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904799.011, "dur": 0.184,
    "args": {
      "Python parent id": 3912, "Python id": 3913, "finished": true, "Ev Idx": 6813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904799.108, "dur": 0.056,
    "args": {
      "Python parent id": 3913, "Python id": 3914, "finished": true, "Ev Idx": 6814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904799.379, "dur": 0.287,
    "args": {
      "Python parent id": 3911, "Python id": 3915, "finished": true, "Ev Idx": 6815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904800.179, "dur": 0.215,
    "args": {
      "Python parent id": 3880, "Python id": 3916, "finished": true, "Ev Idx": 6816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904801.018, "dur": 0.938,
    "args": {
      "Python parent id": 3880, "Python id": 3917, "finished": true, "Ev Idx": 6817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904801.424, "dur": 0.479,
    "args": {
      "Python parent id": 3917, "Python id": 3918, "finished": true, "Ev Idx": 6818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904802.440, "dur": 3.184,
    "args": {
      "Python parent id": 3879, "Python id": 3919, "finished": true, "Ev Idx": 6819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904806.252, "dur": 7.600,
    "args": {
      "Python parent id": 12, "Python id": 3920, "finished": true, "Ev Idx": 6820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904814.332, "dur": 0.096,
    "args": {
      "Python parent id": 12, "Python id": 3921, "finished": true, "Ev Idx": 6821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904815.003, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 3922, "finished": true, "Ev Idx": 6822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904815.663, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 3923, "finished": true, "Ev Idx": 6823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904816.154, "dur": 9.392,
    "args": {
      "Python parent id": 12, "Python id": 3924, "finished": true, "Ev Idx": 6824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904826.689, "dur": 23.223,
    "args": {
      "Python parent id": 12, "Python id": 3925, "finished": true, "Ev Idx": 6825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.031, "dur": 18.683,
    "args": {
      "Python parent id": 3925, "Python id": 3926, "finished": true, "Ev Idx": 6826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.187, "dur": 0.478,
    "args": {
      "Python parent id": 3926, "Python id": 3927, "finished": true, "Ev Idx": 6827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.363, "dur": 0.226,
    "args": {
      "Python parent id": 3927, "Python id": 3928, "finished": true, "Ev Idx": 6828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.459, "dur": 0.095,
    "args": {
      "Python parent id": 3928, "Python id": 3929, "finished": true, "Ev Idx": 6829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.908, "dur": 15.086,
    "args": {
      "Python parent id": 3926, "Python id": 3930, "finished": true, "Ev Idx": 6830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904827.995, "dur": 0.067,
    "args": {
      "Python parent id": 3930, "Python id": 3931, "finished": true, "Ev Idx": 6831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904828.206, "dur": 0.064,
    "args": {
      "Python parent id": 3930, "Python id": 3932, "finished": true, "Ev Idx": 6832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904828.412, "dur": 0.061,
    "args": {
      "Python parent id": 3930, "Python id": 3933, "finished": true, "Ev Idx": 6833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904828.872, "dur": 0.040,
    "args": {
      "Python parent id": 3930, "Python id": 3934, "finished": true, "Ev Idx": 6834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904829.209, "dur": 0.081,
    "args": {
      "Python parent id": 3930, "Python id": 3935, "finished": true, "Ev Idx": 6835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904829.620, "dur": 13.300,
    "args": {
      "Python parent id": 3930, "Python id": 3936, "finished": true, "Ev Idx": 6836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904829.742, "dur": 0.063,
    "args": {
      "Python parent id": 3936, "Python id": 3937, "finished": true, "Ev Idx": 6837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904829.975, "dur": 0.063,
    "args": {
      "Python parent id": 3936, "Python id": 3938, "finished": true, "Ev Idx": 6838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904830.253, "dur": 0.067,
    "args": {
      "Python parent id": 3936, "Python id": 3939, "finished": true, "Ev Idx": 6839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904830.689, "dur": 0.037,
    "args": {
      "Python parent id": 3936, "Python id": 3940, "finished": true, "Ev Idx": 6840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904831.057, "dur": 11.802,
    "args": {
      "Python parent id": 3936, "Python id": 3941, "finished": true, "Ev Idx": 6841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904836.143, "dur": 6.636,
    "args": {
      "Python parent id": 3941, "Python id": 3942, "finished": true, "Ev Idx": 6842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904836.351, "dur": 4.168,
    "args": {
      "Python parent id": 3942, "Python id": 3943, "finished": true, "Ev Idx": 6843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904836.665, "dur": 3.808,
    "args": {
      "Python parent id": 3943, "Python id": 3944, "finished": true, "Ev Idx": 6844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904836.812, "dur": 0.350,
    "args": {
      "Python parent id": 3944, "Python id": 3945, "finished": true, "Ev Idx": 6845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904836.960, "dur": 0.109,
    "args": {
      "Python parent id": 3945, "Python id": 3946, "finished": true, "Ev Idx": 6846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904837.382, "dur": 2.580,
    "args": {
      "Python parent id": 3944, "Python id": 3947, "finished": true, "Ev Idx": 6847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904837.580, "dur": 2.268,
    "args": {
      "Python parent id": 3947, "Python id": 3948, "finished": true, "Ev Idx": 6848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904837.798, "dur": 2.002,
    "args": {
      "Python parent id": 3948, "Python id": 3949, "finished": true, "Ev Idx": 6849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904837.960, "dur": 1.542,
    "args": {
      "Python parent id": 3949, "Python id": 3950, "finished": true, "Ev Idx": 6850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904838.171, "dur": 0.632,
    "args": {
      "Python parent id": 3950, "Python id": 3951, "finished": true, "Ev Idx": 6851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904838.288, "dur": 0.051,
    "args": {
      "Python parent id": 3951, "Python id": 3952, "finished": true, "Ev Idx": 6852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904838.550, "dur": 0.209,
    "args": {
      "Python parent id": 3951, "Python id": 3953, "finished": true, "Ev Idx": 6853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904840.182, "dur": 0.241,
    "args": {
      "Python parent id": 3944, "Python id": 3954, "finished": true, "Ev Idx": 6854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904840.702, "dur": 0.141,
    "args": {
      "Python parent id": 3942, "Python id": 3955, "finished": true, "Ev Idx": 6855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904841.191, "dur": 1.517,
    "args": {
      "Python parent id": 3942, "Python id": 3956, "finished": true, "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904841.352, "dur": 1.308,
    "args": {
      "Python parent id": 3956, "Python id": 3957, "finished": true, "Ev Idx": 6857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904841.556, "dur": 0.372,
    "args": {
      "Python parent id": 3957, "Python id": 3958, "finished": true, "Ev Idx": 6858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904841.690, "dur": 0.194,
    "args": {
      "Python parent id": 3958, "Python id": 3959, "finished": true, "Ev Idx": 6859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904841.778, "dur": 0.058,
    "args": {
      "Python parent id": 3959, "Python id": 3960, "finished": true, "Ev Idx": 6860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904842.067, "dur": 0.561,
    "args": {
      "Python parent id": 3957, "Python id": 3961, "finished": true, "Ev Idx": 6861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904843.122, "dur": 0.297,
    "args": {
      "Python parent id": 3926, "Python id": 3962, "finished": true, "Ev Idx": 6862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904844.209, "dur": 1.360,
    "args": {
      "Python parent id": 3926, "Python id": 3963, "finished": true, "Ev Idx": 6863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904844.734, "dur": 0.725,
    "args": {
      "Python parent id": 3963, "Python id": 3964, "finished": true, "Ev Idx": 6864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904846.036, "dur": 3.811,
    "args": {
      "Python parent id": 3925, "Python id": 3965, "finished": true, "Ev Idx": 6865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740904850.880, "dur": 102.659,
    "args": {
      "Python parent id": 12, "Python id": 3966, "finished": true, "Ev Idx": 6866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904851.360, "dur": 0.475,
    "args": {
      "Python parent id": 3966, "Python id": 3967, "finished": true, "Ev Idx": 6867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904851.664, "dur": 0.103,
    "args": {
      "Python parent id": 3967, "Python id": 3968, "finished": true, "Ev Idx": 6868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904852.148, "dur": 0.523,
    "args": {
      "Python parent id": 3966, "Python id": 3969, "finished": true, "Ev Idx": 6869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740904852.900, "dur": 0.278,
    "args": {
      "Python parent id": 3966, "Python id": 3970, "finished": true, "Ev Idx": 6870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740904853.499, "dur": 3.187,
    "args": {
      "Python parent id": 3966, "Python id": 3971, "finished": true, "Ev Idx": 6871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740904857.411, "dur": 96.027,
    "args": {
      "Python parent id": 3966, "Python id": 3972, "finished": true, "Ev Idx": 6872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740904858.499, "dur": 94.710,
    "args": {
      "Python parent id": 3972, "Python id": 3973, "finished": true, "Ev Idx": 6873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904858.868, "dur": 94.272,
    "args": {
      "Python parent id": 3973, "Python id": 3974, "finished": true, "Ev Idx": 6874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740904866.028, "dur": 85.332,
    "args": {
      "Python parent id": 3974, "Python id": 3975, "finished": true, "Ev Idx": 6875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740904866.348, "dur": 84.922,
    "args": {
      "Python parent id": 3975, "Python id": 3976, "finished": true, "Ev Idx": 6876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740904867.088, "dur": 84.066,
    "args": {
      "Python parent id": 3976, "Python id": 3977, "finished": true, "Ev Idx": 6877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904867.670, "dur": 0.445,
    "args": {
      "Python parent id": 3977, "Python id": 3978, "finished": true, "Ev Idx": 6878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904867.950, "dur": 0.062,
    "args": {
      "Python parent id": 3978, "Python id": 3979, "finished": true, "Ev Idx": 6879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740904868.502, "dur": 0.360,
    "args": {
      "Python parent id": 3977, "Python id": 3980, "finished": true, "Ev Idx": 6880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740904868.753, "dur": 0.058,
    "args": {
      "Python parent id": 3980, "Python id": 3981, "finished": true, "Ev Idx": 6881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904869.209, "dur": 81.797,
    "args": {
      "Python parent id": 3977, "Python id": 3982, "finished": true, "Ev Idx": 6882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740904869.956, "dur": 79.627,
    "args": {
      "Python parent id": 3982, "Python id": 3983, "finished": true, "Ev Idx": 6883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740904878.988, "dur": 21.861,
    "args": {
      "Python parent id": 3983, "Python id": 3984, "finished": true, "Ev Idx": 6884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740904879.123, "dur": 0.146,
    "args": {
      "Python parent id": 3984, "Python id": 3985, "finished": true, "Ev Idx": 6885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740904879.637, "dur": 21.089,
    "args": {
      "Python parent id": 3984, "Python id": 3986, "finished": true, "Ev Idx": 6886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904879.814, "dur": 9.401,
    "args": {
      "Python parent id": 3986, "Python id": 3987, "finished": true, "Ev Idx": 6887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904909.536, "dur": 7.617,
    "args": {
      "Python parent id": 3983, "Python id": 3988, "finished": true, "Ev Idx": 6888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740904926.045, "dur": 8.605,
    "args": {
      "Python parent id": 3983, "Python id": 3989, "finished": true, "Ev Idx": 6889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904954.983, "dur": 18.452,
    "args": {
      "Python parent id": 12, "Python id": 3990, "finished": true, "Ev Idx": 6890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904955.431, "dur": 13.978,
    "args": {
      "Python parent id": 3990, "Python id": 3991, "finished": true, "Ev Idx": 6891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904955.638, "dur": 0.594,
    "args": {
      "Python parent id": 3991, "Python id": 3992, "finished": true, "Ev Idx": 6892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904955.877, "dur": 0.268,
    "args": {
      "Python parent id": 3992, "Python id": 3993, "finished": true, "Ev Idx": 6893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904956.015, "dur": 0.072,
    "args": {
      "Python parent id": 3993, "Python id": 3994, "finished": true, "Ev Idx": 6894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904956.582, "dur": 10.222,
    "args": {
      "Python parent id": 3991, "Python id": 3995, "finished": true, "Ev Idx": 6895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904956.747, "dur": 0.078,
    "args": {
      "Python parent id": 3995, "Python id": 3996, "finished": true, "Ev Idx": 6896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904956.959, "dur": 0.075,
    "args": {
      "Python parent id": 3995, "Python id": 3997, "finished": true, "Ev Idx": 6897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904957.180, "dur": 0.056,
    "args": {
      "Python parent id": 3995, "Python id": 3998, "finished": true, "Ev Idx": 6898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904957.672, "dur": 0.062,
    "args": {
      "Python parent id": 3995, "Python id": 3999, "finished": true, "Ev Idx": 6899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904958.124, "dur": 0.098,
    "args": {
      "Python parent id": 3995, "Python id": 4000, "finished": true, "Ev Idx": 6900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904958.677, "dur": 8.048,
    "args": {
      "Python parent id": 3995, "Python id": 4001, "finished": true, "Ev Idx": 6901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904958.796, "dur": 0.062,
    "args": {
      "Python parent id": 4001, "Python id": 4002, "finished": true, "Ev Idx": 6902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904959.030, "dur": 0.061,
    "args": {
      "Python parent id": 4001, "Python id": 4003, "finished": true, "Ev Idx": 6903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904959.298, "dur": 0.073,
    "args": {
      "Python parent id": 4001, "Python id": 4004, "finished": true, "Ev Idx": 6904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904959.781, "dur": 0.031,
    "args": {
      "Python parent id": 4001, "Python id": 4005, "finished": true, "Ev Idx": 6905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904960.184, "dur": 6.391,
    "args": {
      "Python parent id": 4001, "Python id": 4006, "finished": true, "Ev Idx": 6906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740904960.538, "dur": 5.978,
    "args": {
      "Python parent id": 4006, "Python id": 4007, "finished": true, "Ev Idx": 6907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740904960.750, "dur": 3.910,
    "args": {
      "Python parent id": 4007, "Python id": 4008, "finished": true, "Ev Idx": 6908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740904961.000, "dur": 3.559,
    "args": {
      "Python parent id": 4008, "Python id": 4009, "finished": true, "Ev Idx": 6909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740904961.177, "dur": 0.305,
    "args": {
      "Python parent id": 4009, "Python id": 4010, "finished": true, "Ev Idx": 6910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740904961.307, "dur": 0.071,
    "args": {
      "Python parent id": 4010, "Python id": 4011, "finished": true, "Ev Idx": 6911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740904961.647, "dur": 2.500,
    "args": {
      "Python parent id": 4009, "Python id": 4012, "finished": true, "Ev Idx": 6912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740904961.892, "dur": 2.136,
    "args": {
      "Python parent id": 4012, "Python id": 4013, "finished": true, "Ev Idx": 6913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740904962.109, "dur": 1.869,
    "args": {
      "Python parent id": 4013, "Python id": 4014, "finished": true, "Ev Idx": 6914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740904962.280, "dur": 1.466,
    "args": {
      "Python parent id": 4014, "Python id": 4015, "finished": true, "Ev Idx": 6915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740904962.482, "dur": 0.614,
    "args": {
      "Python parent id": 4015, "Python id": 4016, "finished": true, "Ev Idx": 6916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904962.568, "dur": 0.045,
    "args": {
      "Python parent id": 4016, "Python id": 4017, "finished": true, "Ev Idx": 6917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740904962.817, "dur": 0.207,
    "args": {
      "Python parent id": 4016, "Python id": 4018, "finished": true, "Ev Idx": 6918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740904964.341, "dur": 0.183,
    "args": {
      "Python parent id": 4009, "Python id": 4019, "finished": true, "Ev Idx": 6919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740904964.827, "dur": 0.113,
    "args": {
      "Python parent id": 4007, "Python id": 4020, "finished": true, "Ev Idx": 6920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740904965.212, "dur": 1.262,
    "args": {
      "Python parent id": 4007, "Python id": 4021, "finished": true, "Ev Idx": 6921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740904965.352, "dur": 1.061,
    "args": {
      "Python parent id": 4021, "Python id": 4022, "finished": true, "Ev Idx": 6922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904965.504, "dur": 0.391,
    "args": {
      "Python parent id": 4022, "Python id": 4023, "finished": true, "Ev Idx": 6923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904965.645, "dur": 0.183,
    "args": {
      "Python parent id": 4023, "Python id": 4024, "finished": true, "Ev Idx": 6924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904965.740, "dur": 0.057,
    "args": {
      "Python parent id": 4024, "Python id": 4025, "finished": true, "Ev Idx": 6925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740904966.027, "dur": 0.357,
    "args": {
      "Python parent id": 4022, "Python id": 4026, "finished": true, "Ev Idx": 6926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740904966.973, "dur": 0.312,
    "args": {
      "Python parent id": 3991, "Python id": 4027, "finished": true, "Ev Idx": 6927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740904968.010, "dur": 1.256,
    "args": {
      "Python parent id": 3991, "Python id": 4028, "finished": true, "Ev Idx": 6928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740904968.532, "dur": 0.664,
    "args": {
      "Python parent id": 4028, "Python id": 4029, "finished": true, "Ev Idx": 6929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740904969.723, "dur": 3.639,
    "args": {
      "Python parent id": 3990, "Python id": 4030, "finished": true, "Ev Idx": 6930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740904974.056, "dur": 8.599,
    "args": {
      "Python parent id": 12, "Python id": 4031, "finished": true, "Ev Idx": 6931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740904983.173, "dur": 0.166,
    "args": {
      "Python parent id": 12, "Python id": 4032, "finished": true, "Ev Idx": 6932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904984.281, "dur": 0.221,
    "args": {
      "Python parent id": 12, "Python id": 4033, "finished": true, "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904985.045, "dur": 0.085,
    "args": {
      "Python parent id": 12, "Python id": 4034, "finished": true, "Ev Idx": 6934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740904985.545, "dur": 10.711,
    "args": {
      "Python parent id": 12, "Python id": 4035, "finished": true, "Ev Idx": 6935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740904997.654, "dur": 15.345,
    "args": {
      "Python parent id": 12, "Python id": 4036, "finished": true, "Ev Idx": 6936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740904998.004, "dur": 11.749,
    "args": {
      "Python parent id": 4036, "Python id": 4037, "finished": true, "Ev Idx": 6937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740904998.171, "dur": 0.518,
    "args": {
      "Python parent id": 4037, "Python id": 4038, "finished": true, "Ev Idx": 6938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740904998.340, "dur": 0.273,
    "args": {
      "Python parent id": 4038, "Python id": 4039, "finished": true, "Ev Idx": 6939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740904998.442, "dur": 0.104,
    "args": {
      "Python parent id": 4039, "Python id": 4040, "finished": true, "Ev Idx": 6940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740904999.036, "dur": 8.758,
    "args": {
      "Python parent id": 4037, "Python id": 4041, "finished": true, "Ev Idx": 6941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904999.130, "dur": 0.071,
    "args": {
      "Python parent id": 4041, "Python id": 4042, "finished": true, "Ev Idx": 6942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904999.317, "dur": 0.065,
    "args": {
      "Python parent id": 4041, "Python id": 4043, "finished": true, "Ev Idx": 6943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740904999.518, "dur": 0.062,
    "args": {
      "Python parent id": 4041, "Python id": 4044, "finished": true, "Ev Idx": 6944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740904999.966, "dur": 0.073,
    "args": {
      "Python parent id": 4041, "Python id": 4045, "finished": true, "Ev Idx": 6945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905000.323, "dur": 0.082,
    "args": {
      "Python parent id": 4041, "Python id": 4046, "finished": true, "Ev Idx": 6946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905000.874, "dur": 6.855,
    "args": {
      "Python parent id": 4041, "Python id": 4047, "finished": true, "Ev Idx": 6947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905000.984, "dur": 0.070,
    "args": {
      "Python parent id": 4047, "Python id": 4048, "finished": true, "Ev Idx": 6948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905001.224, "dur": 0.058,
    "args": {
      "Python parent id": 4047, "Python id": 4049, "finished": true, "Ev Idx": 6949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905001.496, "dur": 0.066,
    "args": {
      "Python parent id": 4047, "Python id": 4050, "finished": true, "Ev Idx": 6950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905001.956, "dur": 0.040,
    "args": {
      "Python parent id": 4047, "Python id": 4051, "finished": true, "Ev Idx": 6951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905002.319, "dur": 5.325,
    "args": {
      "Python parent id": 4047, "Python id": 4052, "finished": true, "Ev Idx": 6952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905002.537, "dur": 5.051,
    "args": {
      "Python parent id": 4052, "Python id": 4053, "finished": true, "Ev Idx": 6953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905002.678, "dur": 3.213,
    "args": {
      "Python parent id": 4053, "Python id": 4054, "finished": true, "Ev Idx": 6954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905002.889, "dur": 2.937,
    "args": {
      "Python parent id": 4054, "Python id": 4055, "finished": true, "Ev Idx": 6955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.015, "dur": 0.265,
    "args": {
      "Python parent id": 4055, "Python id": 4056, "finished": true, "Ev Idx": 6956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.140, "dur": 0.078,
    "args": {
      "Python parent id": 4056, "Python id": 4057, "finished": true, "Ev Idx": 6957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.421, "dur": 2.106,
    "args": {
      "Python parent id": 4055, "Python id": 4058, "finished": true, "Ev Idx": 6958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.612, "dur": 1.854,
    "args": {
      "Python parent id": 4058, "Python id": 4059, "finished": true, "Ev Idx": 6959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.832, "dur": 1.586,
    "args": {
      "Python parent id": 4059, "Python id": 4060, "finished": true, "Ev Idx": 6960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905003.977, "dur": 1.190,
    "args": {
      "Python parent id": 4060, "Python id": 4061, "finished": true, "Ev Idx": 6961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905004.143, "dur": 0.501,
    "args": {
      "Python parent id": 4061, "Python id": 4062, "finished": true, "Ev Idx": 6962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905004.225, "dur": 0.047,
    "args": {
      "Python parent id": 4062, "Python id": 4063, "finished": true, "Ev Idx": 6963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905004.434, "dur": 0.156,
    "args": {
      "Python parent id": 4062, "Python id": 4064, "finished": true, "Ev Idx": 6964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905005.711, "dur": 0.076,
    "args": {
      "Python parent id": 4055, "Python id": 4065, "finished": true, "Ev Idx": 6965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.038, "dur": 0.113,
    "args": {
      "Python parent id": 4053, "Python id": 4066, "finished": true, "Ev Idx": 6966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.428, "dur": 1.115,
    "args": {
      "Python parent id": 4053, "Python id": 4067, "finished": true, "Ev Idx": 6967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.574, "dur": 0.925,
    "args": {
      "Python parent id": 4067, "Python id": 4068, "finished": true, "Ev Idx": 6968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.715, "dur": 0.371,
    "args": {
      "Python parent id": 4068, "Python id": 4069, "finished": true, "Ev Idx": 6969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.858, "dur": 0.183,
    "args": {
      "Python parent id": 4069, "Python id": 4070, "finished": true, "Ev Idx": 6970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905006.952, "dur": 0.058,
    "args": {
      "Python parent id": 4070, "Python id": 4071, "finished": true, "Ev Idx": 6971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905007.219, "dur": 0.251,
    "args": {
      "Python parent id": 4068, "Python id": 4072, "finished": true, "Ev Idx": 6972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905007.925, "dur": 0.204,
    "args": {
      "Python parent id": 4037, "Python id": 4073, "finished": true, "Ev Idx": 6973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905008.713, "dur": 0.925,
    "args": {
      "Python parent id": 4037, "Python id": 4074, "finished": true, "Ev Idx": 6974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905009.105, "dur": 0.480,
    "args": {
      "Python parent id": 4074, "Python id": 4075, "finished": true, "Ev Idx": 6975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905010.000, "dur": 2.932,
    "args": {
      "Python parent id": 4036, "Python id": 4076, "finished": true, "Ev Idx": 6976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905013.984, "dur": 93.063,
    "args": {
      "Python parent id": 12, "Python id": 4077, "finished": true, "Ev Idx": 6977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905014.389, "dur": 0.505,
    "args": {
      "Python parent id": 4077, "Python id": 4078, "finished": true, "Ev Idx": 6978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905014.711, "dur": 0.095,
    "args": {
      "Python parent id": 4078, "Python id": 4079, "finished": true, "Ev Idx": 6979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905015.247, "dur": 0.462,
    "args": {
      "Python parent id": 4077, "Python id": 4080, "finished": true, "Ev Idx": 6980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905015.910, "dur": 0.257,
    "args": {
      "Python parent id": 4077, "Python id": 4081, "finished": true, "Ev Idx": 6981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905016.451, "dur": 1.685,
    "args": {
      "Python parent id": 4077, "Python id": 4082, "finished": true, "Ev Idx": 6982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905018.798, "dur": 88.153,
    "args": {
      "Python parent id": 4077, "Python id": 4083, "finished": true, "Ev Idx": 6983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905019.768, "dur": 86.983,
    "args": {
      "Python parent id": 4083, "Python id": 4084, "finished": true, "Ev Idx": 6984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905020.108, "dur": 86.558,
    "args": {
      "Python parent id": 4084, "Python id": 4085, "finished": true, "Ev Idx": 6985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905026.542, "dur": 78.686,
    "args": {
      "Python parent id": 4085, "Python id": 4086, "finished": true, "Ev Idx": 6986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905026.853, "dur": 78.291,
    "args": {
      "Python parent id": 4086, "Python id": 4087, "finished": true, "Ev Idx": 6987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905027.449, "dur": 77.543,
    "args": {
      "Python parent id": 4087, "Python id": 4088, "finished": true, "Ev Idx": 6988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905027.995, "dur": 0.401,
    "args": {
      "Python parent id": 4088, "Python id": 4089, "finished": true, "Ev Idx": 6989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905028.271, "dur": 0.076,
    "args": {
      "Python parent id": 4089, "Python id": 4090, "finished": true, "Ev Idx": 6990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905028.694, "dur": 0.355,
    "args": {
      "Python parent id": 4088, "Python id": 4091, "finished": true, "Ev Idx": 6991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905028.957, "dur": 0.046,
    "args": {
      "Python parent id": 4091, "Python id": 4092, "finished": true, "Ev Idx": 6992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905029.365, "dur": 75.539,
    "args": {
      "Python parent id": 4088, "Python id": 4093, "finished": true, "Ev Idx": 6993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905030.046, "dur": 73.449,
    "args": {
      "Python parent id": 4093, "Python id": 4094, "finished": true, "Ev Idx": 6994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905037.981, "dur": 18.088,
    "args": {
      "Python parent id": 4094, "Python id": 4095, "finished": true, "Ev Idx": 6995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905038.112, "dur": 0.160,
    "args": {
      "Python parent id": 4095, "Python id": 4096, "finished": true, "Ev Idx": 6996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905038.638, "dur": 17.242,
    "args": {
      "Python parent id": 4095, "Python id": 4097, "finished": true, "Ev Idx": 6997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905038.772, "dur": 7.435,
    "args": {
      "Python parent id": 4097, "Python id": 4098, "finished": true, "Ev Idx": 6998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905064.329, "dur": 8.257,
    "args": {
      "Python parent id": 4094, "Python id": 4099, "finished": true, "Ev Idx": 6999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905080.892, "dur": 8.169,
    "args": {
      "Python parent id": 4094, "Python id": 4100, "finished": true, "Ev Idx": 7000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905108.552, "dur": 18.998,
    "args": {
      "Python parent id": 12, "Python id": 4101, "finished": true, "Ev Idx": 7001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905108.926, "dur": 14.778,
    "args": {
      "Python parent id": 4101, "Python id": 4102, "finished": true, "Ev Idx": 7002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905109.125, "dur": 0.503,
    "args": {
      "Python parent id": 4102, "Python id": 4103, "finished": true, "Ev Idx": 7003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905109.325, "dur": 0.241,
    "args": {
      "Python parent id": 4103, "Python id": 4104, "finished": true, "Ev Idx": 7004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905109.470, "dur": 0.062,
    "args": {
      "Python parent id": 4104, "Python id": 4105, "finished": true, "Ev Idx": 7005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905110.005, "dur": 11.280,
    "args": {
      "Python parent id": 4102, "Python id": 4106, "finished": true, "Ev Idx": 7006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905110.104, "dur": 0.067,
    "args": {
      "Python parent id": 4106, "Python id": 4107, "finished": true, "Ev Idx": 7007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905110.276, "dur": 0.061,
    "args": {
      "Python parent id": 4106, "Python id": 4108, "finished": true, "Ev Idx": 7008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905110.496, "dur": 0.059,
    "args": {
      "Python parent id": 4106, "Python id": 4109, "finished": true, "Ev Idx": 7009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905111.070, "dur": 0.040,
    "args": {
      "Python parent id": 4106, "Python id": 4110, "finished": true, "Ev Idx": 7010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905111.462, "dur": 0.084,
    "args": {
      "Python parent id": 4106, "Python id": 4111, "finished": true, "Ev Idx": 7011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905111.968, "dur": 9.216,
    "args": {
      "Python parent id": 4106, "Python id": 4112, "finished": true, "Ev Idx": 7012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905112.062, "dur": 0.065,
    "args": {
      "Python parent id": 4112, "Python id": 4113, "finished": true, "Ev Idx": 7013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905112.279, "dur": 1.537,
    "args": {
      "Python parent id": 4112, "Python id": 4114, "finished": true, "Ev Idx": 7014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905114.072, "dur": 0.059,
    "args": {
      "Python parent id": 4112, "Python id": 4115, "finished": true, "Ev Idx": 7015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905114.491, "dur": 0.039,
    "args": {
      "Python parent id": 4112, "Python id": 4116, "finished": true, "Ev Idx": 7016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905114.898, "dur": 6.153,
    "args": {
      "Python parent id": 4112, "Python id": 4117, "finished": true, "Ev Idx": 7017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905115.214, "dur": 5.761,
    "args": {
      "Python parent id": 4117, "Python id": 4118, "finished": true, "Ev Idx": 7018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905115.411, "dur": 3.709,
    "args": {
      "Python parent id": 4118, "Python id": 4119, "finished": true, "Ev Idx": 7019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905115.625, "dur": 3.433,
    "args": {
      "Python parent id": 4119, "Python id": 4120, "finished": true, "Ev Idx": 7020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905115.774, "dur": 0.296,
    "args": {
      "Python parent id": 4120, "Python id": 4121, "finished": true, "Ev Idx": 7021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905115.914, "dur": 0.073,
    "args": {
      "Python parent id": 4121, "Python id": 4122, "finished": true, "Ev Idx": 7022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905116.287, "dur": 2.357,
    "args": {
      "Python parent id": 4120, "Python id": 4123, "finished": true, "Ev Idx": 7023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905116.503, "dur": 2.066,
    "args": {
      "Python parent id": 4123, "Python id": 4124, "finished": true, "Ev Idx": 7024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905116.720, "dur": 1.778,
    "args": {
      "Python parent id": 4124, "Python id": 4125, "finished": true, "Ev Idx": 7025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905116.890, "dur": 1.288,
    "args": {
      "Python parent id": 4125, "Python id": 4126, "finished": true, "Ev Idx": 7026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905117.122, "dur": 0.538,
    "args": {
      "Python parent id": 4126, "Python id": 4127, "finished": true, "Ev Idx": 7027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905117.223, "dur": 0.046,
    "args": {
      "Python parent id": 4127, "Python id": 4128, "finished": true, "Ev Idx": 7028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905117.446, "dur": 0.158,
    "args": {
      "Python parent id": 4127, "Python id": 4129, "finished": true, "Ev Idx": 7029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905118.835, "dur": 0.161,
    "args": {
      "Python parent id": 4120, "Python id": 4130, "finished": true, "Ev Idx": 7030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905119.308, "dur": 0.078,
    "args": {
      "Python parent id": 4118, "Python id": 4131, "finished": true, "Ev Idx": 7031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905119.720, "dur": 1.214,
    "args": {
      "Python parent id": 4118, "Python id": 4132, "finished": true, "Ev Idx": 7032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905119.874, "dur": 0.998,
    "args": {
      "Python parent id": 4132, "Python id": 4133, "finished": true, "Ev Idx": 7033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905120.019, "dur": 0.382,
    "args": {
      "Python parent id": 4133, "Python id": 4134, "finished": true, "Ev Idx": 7034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905120.159, "dur": 0.176,
    "args": {
      "Python parent id": 4134, "Python id": 4135, "finished": true, "Ev Idx": 7035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905120.246, "dur": 0.058,
    "args": {
      "Python parent id": 4135, "Python id": 4136, "finished": true, "Ev Idx": 7036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905120.535, "dur": 0.272,
    "args": {
      "Python parent id": 4133, "Python id": 4137, "finished": true, "Ev Idx": 7037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905121.413, "dur": 0.295,
    "args": {
      "Python parent id": 4102, "Python id": 4138, "finished": true, "Ev Idx": 7038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905122.336, "dur": 1.210,
    "args": {
      "Python parent id": 4102, "Python id": 4139, "finished": true, "Ev Idx": 7039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905122.895, "dur": 0.581,
    "args": {
      "Python parent id": 4139, "Python id": 4140, "finished": true, "Ev Idx": 7040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905124.028, "dur": 3.447,
    "args": {
      "Python parent id": 4101, "Python id": 4141, "finished": true, "Ev Idx": 7041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905128.179, "dur": 7.459,
    "args": {
      "Python parent id": 12, "Python id": 4142, "finished": true, "Ev Idx": 7042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905136.186, "dur": 0.112,
    "args": {
      "Python parent id": 12, "Python id": 4143, "finished": true, "Ev Idx": 7043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905136.963, "dur": 0.220,
    "args": {
      "Python parent id": 12, "Python id": 4144, "finished": true, "Ev Idx": 7044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905137.638, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 4145, "finished": true, "Ev Idx": 7045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905138.125, "dur": 9.629,
    "args": {
      "Python parent id": 12, "Python id": 4146, "finished": true, "Ev Idx": 7046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905149.173, "dur": 14.927,
    "args": {
      "Python parent id": 12, "Python id": 4147, "finished": true, "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905149.561, "dur": 11.507,
    "args": {
      "Python parent id": 4147, "Python id": 4148, "finished": true, "Ev Idx": 7048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905149.752, "dur": 0.462,
    "args": {
      "Python parent id": 4148, "Python id": 4149, "finished": true, "Ev Idx": 7049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905149.906, "dur": 0.232,
    "args": {
      "Python parent id": 4149, "Python id": 4150, "finished": true, "Ev Idx": 7050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905150.006, "dur": 0.096,
    "args": {
      "Python parent id": 4150, "Python id": 4151, "finished": true, "Ev Idx": 7051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905150.532, "dur": 8.617,
    "args": {
      "Python parent id": 4148, "Python id": 4152, "finished": true, "Ev Idx": 7052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905150.620, "dur": 0.075,
    "args": {
      "Python parent id": 4152, "Python id": 4153, "finished": true, "Ev Idx": 7053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905150.814, "dur": 0.060,
    "args": {
      "Python parent id": 4152, "Python id": 4154, "finished": true, "Ev Idx": 7054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905151.018, "dur": 0.058,
    "args": {
      "Python parent id": 4152, "Python id": 4155, "finished": true, "Ev Idx": 7055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905151.492, "dur": 0.070,
    "args": {
      "Python parent id": 4152, "Python id": 4156, "finished": true, "Ev Idx": 7056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905151.847, "dur": 0.077,
    "args": {
      "Python parent id": 4152, "Python id": 4157, "finished": true, "Ev Idx": 7057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905152.258, "dur": 6.833,
    "args": {
      "Python parent id": 4152, "Python id": 4158, "finished": true, "Ev Idx": 7058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905152.423, "dur": 0.060,
    "args": {
      "Python parent id": 4158, "Python id": 4159, "finished": true, "Ev Idx": 7059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905152.630, "dur": 0.057,
    "args": {
      "Python parent id": 4158, "Python id": 4160, "finished": true, "Ev Idx": 7060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905152.890, "dur": 0.074,
    "args": {
      "Python parent id": 4158, "Python id": 4161, "finished": true, "Ev Idx": 7061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905153.309, "dur": 0.045,
    "args": {
      "Python parent id": 4158, "Python id": 4162, "finished": true, "Ev Idx": 7062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905153.704, "dur": 5.304,
    "args": {
      "Python parent id": 4158, "Python id": 4163, "finished": true, "Ev Idx": 7063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905154.000, "dur": 4.950,
    "args": {
      "Python parent id": 4163, "Python id": 4164, "finished": true, "Ev Idx": 7064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905154.180, "dur": 3.030,
    "args": {
      "Python parent id": 4164, "Python id": 4165, "finished": true, "Ev Idx": 7065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905154.382, "dur": 2.785,
    "args": {
      "Python parent id": 4165, "Python id": 4166, "finished": true, "Ev Idx": 7066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905154.523, "dur": 0.292,
    "args": {
      "Python parent id": 4166, "Python id": 4167, "finished": true, "Ev Idx": 7067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905154.692, "dur": 0.064,
    "args": {
      "Python parent id": 4167, "Python id": 4168, "finished": true, "Ev Idx": 7068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.016, "dur": 1.860,
    "args": {
      "Python parent id": 4166, "Python id": 4169, "finished": true, "Ev Idx": 7069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.218, "dur": 1.598,
    "args": {
      "Python parent id": 4169, "Python id": 4170, "finished": true, "Ev Idx": 7070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.401, "dur": 1.370,
    "args": {
      "Python parent id": 4170, "Python id": 4171, "finished": true, "Ev Idx": 7071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.555, "dur": 0.999,
    "args": {
      "Python parent id": 4171, "Python id": 4172, "finished": true, "Ev Idx": 7072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.745, "dur": 0.463,
    "args": {
      "Python parent id": 4172, "Python id": 4173, "finished": true, "Ev Idx": 7073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905155.829, "dur": 0.051,
    "args": {
      "Python parent id": 4173, "Python id": 4174, "finished": true, "Ev Idx": 7074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905156.030, "dur": 0.134,
    "args": {
      "Python parent id": 4173, "Python id": 4175, "finished": true, "Ev Idx": 7075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905157.066, "dur": 0.067,
    "args": {
      "Python parent id": 4166, "Python id": 4176, "finished": true, "Ev Idx": 7076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905157.375, "dur": 0.099,
    "args": {
      "Python parent id": 4164, "Python id": 4177, "finished": true, "Ev Idx": 7077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905157.743, "dur": 1.164,
    "args": {
      "Python parent id": 4164, "Python id": 4178, "finished": true, "Ev Idx": 7078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905157.910, "dur": 0.954,
    "args": {
      "Python parent id": 4178, "Python id": 4179, "finished": true, "Ev Idx": 7079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905158.055, "dur": 0.387,
    "args": {
      "Python parent id": 4179, "Python id": 4180, "finished": true, "Ev Idx": 7080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905158.222, "dur": 0.170,
    "args": {
      "Python parent id": 4180, "Python id": 4181, "finished": true, "Ev Idx": 7081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905158.300, "dur": 0.062,
    "args": {
      "Python parent id": 4181, "Python id": 4182, "finished": true, "Ev Idx": 7082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905158.575, "dur": 0.256,
    "args": {
      "Python parent id": 4179, "Python id": 4183, "finished": true, "Ev Idx": 7083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905159.272, "dur": 0.165,
    "args": {
      "Python parent id": 4148, "Python id": 4184, "finished": true, "Ev Idx": 7084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905160.008, "dur": 0.937,
    "args": {
      "Python parent id": 4148, "Python id": 4185, "finished": true, "Ev Idx": 7085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905160.421, "dur": 0.451,
    "args": {
      "Python parent id": 4185, "Python id": 4186, "finished": true, "Ev Idx": 7086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905161.312, "dur": 2.739,
    "args": {
      "Python parent id": 4147, "Python id": 4187, "finished": true, "Ev Idx": 7087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905165.095, "dur": 96.405,
    "args": {
      "Python parent id": 12, "Python id": 4188, "finished": true, "Ev Idx": 7088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905165.461, "dur": 0.455,
    "args": {
      "Python parent id": 4188, "Python id": 4189, "finished": true, "Ev Idx": 7089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905165.747, "dur": 0.121,
    "args": {
      "Python parent id": 4189, "Python id": 4190, "finished": true, "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905166.219, "dur": 0.392,
    "args": {
      "Python parent id": 4188, "Python id": 4191, "finished": true, "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905166.822, "dur": 0.265,
    "args": {
      "Python parent id": 4188, "Python id": 4192, "finished": true, "Ev Idx": 7092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905167.321, "dur": 1.637,
    "args": {
      "Python parent id": 4188, "Python id": 4193, "finished": true, "Ev Idx": 7093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905169.531, "dur": 91.840,
    "args": {
      "Python parent id": 4188, "Python id": 4194, "finished": true, "Ev Idx": 7094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905170.327, "dur": 90.864,
    "args": {
      "Python parent id": 4194, "Python id": 4195, "finished": true, "Ev Idx": 7095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905170.623, "dur": 90.510,
    "args": {
      "Python parent id": 4195, "Python id": 4196, "finished": true, "Ev Idx": 7096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905176.648, "dur": 83.177,
    "args": {
      "Python parent id": 4196, "Python id": 4197, "finished": true, "Ev Idx": 7097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905176.954, "dur": 82.771,
    "args": {
      "Python parent id": 4197, "Python id": 4198, "finished": true, "Ev Idx": 7098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905177.524, "dur": 82.130,
    "args": {
      "Python parent id": 4198, "Python id": 4199, "finished": true, "Ev Idx": 7099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905178.095, "dur": 0.453,
    "args": {
      "Python parent id": 4199, "Python id": 4200, "finished": true, "Ev Idx": 7100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905178.401, "dur": 0.088,
    "args": {
      "Python parent id": 4200, "Python id": 4201, "finished": true, "Ev Idx": 7101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905178.903, "dur": 0.329,
    "args": {
      "Python parent id": 4199, "Python id": 4202, "finished": true, "Ev Idx": 7102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905179.124, "dur": 0.057,
    "args": {
      "Python parent id": 4202, "Python id": 4203, "finished": true, "Ev Idx": 7103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905179.629, "dur": 79.911,
    "args": {
      "Python parent id": 4199, "Python id": 4204, "finished": true, "Ev Idx": 7104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905180.359, "dur": 77.941,
    "args": {
      "Python parent id": 4204, "Python id": 4205, "finished": true, "Ev Idx": 7105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905187.820, "dur": 18.042,
    "args": {
      "Python parent id": 4205, "Python id": 4206, "finished": true, "Ev Idx": 7106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905187.961, "dur": 0.143,
    "args": {
      "Python parent id": 4206, "Python id": 4207, "finished": true, "Ev Idx": 7107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905188.418, "dur": 17.284,
    "args": {
      "Python parent id": 4206, "Python id": 4208, "finished": true, "Ev Idx": 7108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905188.587, "dur": 7.934,
    "args": {
      "Python parent id": 4208, "Python id": 4209, "finished": true, "Ev Idx": 7109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905213.279, "dur": 6.569,
    "args": {
      "Python parent id": 4205, "Python id": 4210, "finished": true, "Ev Idx": 7110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905233.573, "dur": 9.515,
    "args": {
      "Python parent id": 4205, "Python id": 4211, "finished": true, "Ev Idx": 7111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905262.857, "dur": 17.371,
    "args": {
      "Python parent id": 12, "Python id": 4212, "finished": true, "Ev Idx": 7112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905263.203, "dur": 13.351,
    "args": {
      "Python parent id": 4212, "Python id": 4213, "finished": true, "Ev Idx": 7113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905263.417, "dur": 0.554,
    "args": {
      "Python parent id": 4213, "Python id": 4214, "finished": true, "Ev Idx": 7114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905263.621, "dur": 0.287,
    "args": {
      "Python parent id": 4214, "Python id": 4215, "finished": true, "Ev Idx": 7115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905263.704, "dur": 0.149,
    "args": {
      "Python parent id": 4215, "Python id": 4216, "finished": true, "Ev Idx": 7116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905264.330, "dur": 9.816,
    "args": {
      "Python parent id": 4213, "Python id": 4217, "finished": true, "Ev Idx": 7117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905264.421, "dur": 0.087,
    "args": {
      "Python parent id": 4217, "Python id": 4218, "finished": true, "Ev Idx": 7118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905264.612, "dur": 0.061,
    "args": {
      "Python parent id": 4217, "Python id": 4219, "finished": true, "Ev Idx": 7119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905264.837, "dur": 0.062,
    "args": {
      "Python parent id": 4217, "Python id": 4220, "finished": true, "Ev Idx": 7120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905265.394, "dur": 0.038,
    "args": {
      "Python parent id": 4217, "Python id": 4221, "finished": true, "Ev Idx": 7121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905265.801, "dur": 0.090,
    "args": {
      "Python parent id": 4217, "Python id": 4222, "finished": true, "Ev Idx": 7122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905266.320, "dur": 7.772,
    "args": {
      "Python parent id": 4217, "Python id": 4223, "finished": true, "Ev Idx": 7123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905266.471, "dur": 0.063,
    "args": {
      "Python parent id": 4223, "Python id": 4224, "finished": true, "Ev Idx": 7124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905266.711, "dur": 0.063,
    "args": {
      "Python parent id": 4223, "Python id": 4225, "finished": true, "Ev Idx": 7125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905266.966, "dur": 0.072,
    "args": {
      "Python parent id": 4223, "Python id": 4226, "finished": true, "Ev Idx": 7126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905267.439, "dur": 0.035,
    "args": {
      "Python parent id": 4223, "Python id": 4227, "finished": true, "Ev Idx": 7127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905267.786, "dur": 6.161,
    "args": {
      "Python parent id": 4223, "Python id": 4228, "finished": true, "Ev Idx": 7128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905268.072, "dur": 5.810,
    "args": {
      "Python parent id": 4228, "Python id": 4229, "finished": true, "Ev Idx": 7129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905268.216, "dur": 3.681,
    "args": {
      "Python parent id": 4229, "Python id": 4230, "finished": true, "Ev Idx": 7130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905268.416, "dur": 3.392,
    "args": {
      "Python parent id": 4230, "Python id": 4231, "finished": true, "Ev Idx": 7131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905268.546, "dur": 0.299,
    "args": {
      "Python parent id": 4231, "Python id": 4232, "finished": true, "Ev Idx": 7132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905268.702, "dur": 0.076,
    "args": {
      "Python parent id": 4232, "Python id": 4233, "finished": true, "Ev Idx": 7133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.012, "dur": 2.406,
    "args": {
      "Python parent id": 4231, "Python id": 4234, "finished": true, "Ev Idx": 7134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.249, "dur": 2.019,
    "args": {
      "Python parent id": 4234, "Python id": 4235, "finished": true, "Ev Idx": 7135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.451, "dur": 1.733,
    "args": {
      "Python parent id": 4235, "Python id": 4236, "finished": true, "Ev Idx": 7136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.617, "dur": 1.356,
    "args": {
      "Python parent id": 4236, "Python id": 4237, "finished": true, "Ev Idx": 7137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.816, "dur": 0.612,
    "args": {
      "Python parent id": 4237, "Python id": 4238, "finished": true, "Ev Idx": 7138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905269.921, "dur": 0.064,
    "args": {
      "Python parent id": 4238, "Python id": 4239, "finished": true, "Ev Idx": 7139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905270.186, "dur": 0.174,
    "args": {
      "Python parent id": 4238, "Python id": 4240, "finished": true, "Ev Idx": 7140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905271.624, "dur": 0.149,
    "args": {
      "Python parent id": 4231, "Python id": 4241, "finished": true, "Ev Idx": 7141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905272.088, "dur": 0.164,
    "args": {
      "Python parent id": 4229, "Python id": 4242, "finished": true, "Ev Idx": 7142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905272.533, "dur": 1.300,
    "args": {
      "Python parent id": 4229, "Python id": 4243, "finished": true, "Ev Idx": 7143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905272.711, "dur": 1.082,
    "args": {
      "Python parent id": 4243, "Python id": 4244, "finished": true, "Ev Idx": 7144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905272.872, "dur": 0.406,
    "args": {
      "Python parent id": 4244, "Python id": 4245, "finished": true, "Ev Idx": 7145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905273.042, "dur": 0.176,
    "args": {
      "Python parent id": 4245, "Python id": 4246, "finished": true, "Ev Idx": 7146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905273.126, "dur": 0.057,
    "args": {
      "Python parent id": 4246, "Python id": 4247, "finished": true, "Ev Idx": 7147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905273.436, "dur": 0.301,
    "args": {
      "Python parent id": 4244, "Python id": 4248, "finished": true, "Ev Idx": 7148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905274.307, "dur": 0.251,
    "args": {
      "Python parent id": 4213, "Python id": 4249, "finished": true, "Ev Idx": 7149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905275.204, "dur": 1.212,
    "args": {
      "Python parent id": 4213, "Python id": 4250, "finished": true, "Ev Idx": 7150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905275.769, "dur": 0.578,
    "args": {
      "Python parent id": 4250, "Python id": 4251, "finished": true, "Ev Idx": 7151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905276.937, "dur": 3.226,
    "args": {
      "Python parent id": 4212, "Python id": 4252, "finished": true, "Ev Idx": 7152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905280.826, "dur": 8.551,
    "args": {
      "Python parent id": 12, "Python id": 4253, "finished": true, "Ev Idx": 7153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905289.798, "dur": 0.104,
    "args": {
      "Python parent id": 12, "Python id": 4254, "finished": true, "Ev Idx": 7154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905290.554, "dur": 0.171,
    "args": {
      "Python parent id": 12, "Python id": 4255, "finished": true, "Ev Idx": 7155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905291.193, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 4256, "finished": true, "Ev Idx": 7156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905291.622, "dur": 8.989,
    "args": {
      "Python parent id": 12, "Python id": 4257, "finished": true, "Ev Idx": 7157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905301.794, "dur": 14.208,
    "args": {
      "Python parent id": 12, "Python id": 4258, "finished": true, "Ev Idx": 7158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905302.132, "dur": 10.893,
    "args": {
      "Python parent id": 4258, "Python id": 4259, "finished": true, "Ev Idx": 7159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905302.271, "dur": 0.453,
    "args": {
      "Python parent id": 4259, "Python id": 4260, "finished": true, "Ev Idx": 7160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905302.415, "dur": 0.214,
    "args": {
      "Python parent id": 4260, "Python id": 4261, "finished": true, "Ev Idx": 7161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905302.508, "dur": 0.087,
    "args": {
      "Python parent id": 4261, "Python id": 4262, "finished": true, "Ev Idx": 7162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905302.999, "dur": 8.184,
    "args": {
      "Python parent id": 4259, "Python id": 4263, "finished": true, "Ev Idx": 7163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905303.102, "dur": 0.120,
    "args": {
      "Python parent id": 4263, "Python id": 4264, "finished": true, "Ev Idx": 7164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905303.318, "dur": 0.057,
    "args": {
      "Python parent id": 4263, "Python id": 4265, "finished": true, "Ev Idx": 7165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905303.509, "dur": 0.055,
    "args": {
      "Python parent id": 4263, "Python id": 4266, "finished": true, "Ev Idx": 7166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905303.989, "dur": 0.041,
    "args": {
      "Python parent id": 4263, "Python id": 4267, "finished": true, "Ev Idx": 7167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905304.299, "dur": 0.073,
    "args": {
      "Python parent id": 4263, "Python id": 4268, "finished": true, "Ev Idx": 7168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905304.698, "dur": 6.433,
    "args": {
      "Python parent id": 4263, "Python id": 4269, "finished": true, "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905304.790, "dur": 0.058,
    "args": {
      "Python parent id": 4269, "Python id": 4270, "finished": true, "Ev Idx": 7170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905304.991, "dur": 0.057,
    "args": {
      "Python parent id": 4269, "Python id": 4271, "finished": true, "Ev Idx": 7171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905305.220, "dur": 0.064,
    "args": {
      "Python parent id": 4269, "Python id": 4272, "finished": true, "Ev Idx": 7172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905305.633, "dur": 0.040,
    "args": {
      "Python parent id": 4269, "Python id": 4273, "finished": true, "Ev Idx": 7173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905305.952, "dur": 5.079,
    "args": {
      "Python parent id": 4269, "Python id": 4274, "finished": true, "Ev Idx": 7174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905306.191, "dur": 4.768,
    "args": {
      "Python parent id": 4274, "Python id": 4275, "finished": true, "Ev Idx": 7175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905306.351, "dur": 2.968,
    "args": {
      "Python parent id": 4275, "Python id": 4276, "finished": true, "Ev Idx": 7176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905306.546, "dur": 2.730,
    "args": {
      "Python parent id": 4276, "Python id": 4277, "finished": true, "Ev Idx": 7177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905306.674, "dur": 0.294,
    "args": {
      "Python parent id": 4277, "Python id": 4278, "finished": true, "Ev Idx": 7178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905306.807, "dur": 0.066,
    "args": {
      "Python parent id": 4278, "Python id": 4279, "finished": true, "Ev Idx": 7179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.122, "dur": 1.869,
    "args": {
      "Python parent id": 4277, "Python id": 4280, "finished": true, "Ev Idx": 7180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.311, "dur": 1.610,
    "args": {
      "Python parent id": 4280, "Python id": 4281, "finished": true, "Ev Idx": 7181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.444, "dur": 1.384,
    "args": {
      "Python parent id": 4281, "Python id": 4282, "finished": true, "Ev Idx": 7182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.561, "dur": 1.008,
    "args": {
      "Python parent id": 4282, "Python id": 4283, "finished": true, "Ev Idx": 7183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.718, "dur": 0.444,
    "args": {
      "Python parent id": 4283, "Python id": 4284, "finished": true, "Ev Idx": 7184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.799, "dur": 0.043,
    "args": {
      "Python parent id": 4284, "Python id": 4285, "finished": true, "Ev Idx": 7185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905307.992, "dur": 0.109,
    "args": {
      "Python parent id": 4284, "Python id": 4286, "finished": true, "Ev Idx": 7186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905309.180, "dur": 0.061,
    "args": {
      "Python parent id": 4277, "Python id": 4287, "finished": true, "Ev Idx": 7187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905309.499, "dur": 0.080,
    "args": {
      "Python parent id": 4275, "Python id": 4288, "finished": true, "Ev Idx": 7188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905309.874, "dur": 1.044,
    "args": {
      "Python parent id": 4275, "Python id": 4289, "finished": true, "Ev Idx": 7189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905310.019, "dur": 0.857,
    "args": {
      "Python parent id": 4289, "Python id": 4290, "finished": true, "Ev Idx": 7190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905310.152, "dur": 0.364,
    "args": {
      "Python parent id": 4290, "Python id": 4291, "finished": true, "Ev Idx": 7191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905310.278, "dur": 0.175,
    "args": {
      "Python parent id": 4291, "Python id": 4292, "finished": true, "Ev Idx": 7192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905310.363, "dur": 0.059,
    "args": {
      "Python parent id": 4292, "Python id": 4293, "finished": true, "Ev Idx": 7193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905310.637, "dur": 0.208,
    "args": {
      "Python parent id": 4290, "Python id": 4294, "finished": true, "Ev Idx": 7194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905311.318, "dur": 0.152,
    "args": {
      "Python parent id": 4259, "Python id": 4295, "finished": true, "Ev Idx": 7195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905312.093, "dur": 0.820,
    "args": {
      "Python parent id": 4259, "Python id": 4296, "finished": true, "Ev Idx": 7196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905312.475, "dur": 0.393,
    "args": {
      "Python parent id": 4296, "Python id": 4297, "finished": true, "Ev Idx": 7197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905313.344, "dur": 2.617,
    "args": {
      "Python parent id": 4258, "Python id": 4298, "finished": true, "Ev Idx": 7198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905316.925, "dur": 84.976,
    "args": {
      "Python parent id": 12, "Python id": 4299, "finished": true, "Ev Idx": 7199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905317.283, "dur": 0.406,
    "args": {
      "Python parent id": 4299, "Python id": 4300, "finished": true, "Ev Idx": 7200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905317.546, "dur": 0.098,
    "args": {
      "Python parent id": 4300, "Python id": 4301, "finished": true, "Ev Idx": 7201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905317.961, "dur": 0.395,
    "args": {
      "Python parent id": 4299, "Python id": 4302, "finished": true, "Ev Idx": 7202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905318.554, "dur": 0.237,
    "args": {
      "Python parent id": 4299, "Python id": 4303, "finished": true, "Ev Idx": 7203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905319.032, "dur": 1.535,
    "args": {
      "Python parent id": 4299, "Python id": 4304, "finished": true, "Ev Idx": 7204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905321.072, "dur": 80.706,
    "args": {
      "Python parent id": 4299, "Python id": 4305, "finished": true, "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905321.773, "dur": 79.869,
    "args": {
      "Python parent id": 4305, "Python id": 4306, "finished": true, "Ev Idx": 7206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905322.043, "dur": 79.521,
    "args": {
      "Python parent id": 4306, "Python id": 4307, "finished": true, "Ev Idx": 7207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905327.101, "dur": 73.303,
    "args": {
      "Python parent id": 4307, "Python id": 4308, "finished": true, "Ev Idx": 7208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905327.395, "dur": 72.949,
    "args": {
      "Python parent id": 4308, "Python id": 4309, "finished": true, "Ev Idx": 7209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905327.971, "dur": 72.249,
    "args": {
      "Python parent id": 4309, "Python id": 4310, "finished": true, "Ev Idx": 7210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905328.454, "dur": 0.399,
    "args": {
      "Python parent id": 4310, "Python id": 4311, "finished": true, "Ev Idx": 7211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905328.734, "dur": 0.044,
    "args": {
      "Python parent id": 4311, "Python id": 4312, "finished": true, "Ev Idx": 7212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905329.253, "dur": 0.339,
    "args": {
      "Python parent id": 4310, "Python id": 4313, "finished": true, "Ev Idx": 7213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905329.503, "dur": 0.045,
    "args": {
      "Python parent id": 4313, "Python id": 4314, "finished": true, "Ev Idx": 7214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905329.943, "dur": 70.168,
    "args": {
      "Python parent id": 4310, "Python id": 4315, "finished": true, "Ev Idx": 7215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905330.517, "dur": 68.481,
    "args": {
      "Python parent id": 4315, "Python id": 4316, "finished": true, "Ev Idx": 7216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905337.040, "dur": 16.818,
    "args": {
      "Python parent id": 4316, "Python id": 4317, "finished": true, "Ev Idx": 7217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905337.169, "dur": 0.129,
    "args": {
      "Python parent id": 4317, "Python id": 4318, "finished": true, "Ev Idx": 7218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905337.575, "dur": 16.227,
    "args": {
      "Python parent id": 4317, "Python id": 4319, "finished": true, "Ev Idx": 7219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905337.706, "dur": 6.988,
    "args": {
      "Python parent id": 4319, "Python id": 4320, "finished": true, "Ev Idx": 7220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905361.330, "dur": 6.359,
    "args": {
      "Python parent id": 4316, "Python id": 4321, "finished": true, "Ev Idx": 7221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905376.795, "dur": 8.169,
    "args": {
      "Python parent id": 4316, "Python id": 4322, "finished": true, "Ev Idx": 7222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905402.984, "dur": 16.097,
    "args": {
      "Python parent id": 12, "Python id": 4323, "finished": true, "Ev Idx": 7223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905403.289, "dur": 11.910,
    "args": {
      "Python parent id": 4323, "Python id": 4324, "finished": true, "Ev Idx": 7224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905403.490, "dur": 0.444,
    "args": {
      "Python parent id": 4324, "Python id": 4325, "finished": true, "Ev Idx": 7225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905403.653, "dur": 0.221,
    "args": {
      "Python parent id": 4325, "Python id": 4326, "finished": true, "Ev Idx": 7226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905403.761, "dur": 0.081,
    "args": {
      "Python parent id": 4326, "Python id": 4327, "finished": true, "Ev Idx": 7227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905404.253, "dur": 8.796,
    "args": {
      "Python parent id": 4324, "Python id": 4328, "finished": true, "Ev Idx": 7228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905404.357, "dur": 0.064,
    "args": {
      "Python parent id": 4328, "Python id": 4329, "finished": true, "Ev Idx": 7229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905404.547, "dur": 0.074,
    "args": {
      "Python parent id": 4328, "Python id": 4330, "finished": true, "Ev Idx": 7230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905404.784, "dur": 0.054,
    "args": {
      "Python parent id": 4328, "Python id": 4331, "finished": true, "Ev Idx": 7231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905405.318, "dur": 0.048,
    "args": {
      "Python parent id": 4328, "Python id": 4332, "finished": true, "Ev Idx": 7232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905405.724, "dur": 0.081,
    "args": {
      "Python parent id": 4328, "Python id": 4333, "finished": true, "Ev Idx": 7233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905406.158, "dur": 6.822,
    "args": {
      "Python parent id": 4328, "Python id": 4334, "finished": true, "Ev Idx": 7234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905406.253, "dur": 0.060,
    "args": {
      "Python parent id": 4334, "Python id": 4335, "finished": true, "Ev Idx": 7235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905406.448, "dur": 0.054,
    "args": {
      "Python parent id": 4334, "Python id": 4336, "finished": true, "Ev Idx": 7236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905406.676, "dur": 0.058,
    "args": {
      "Python parent id": 4334, "Python id": 4337, "finished": true, "Ev Idx": 7237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905406.994, "dur": 0.051,
    "args": {
      "Python parent id": 4334, "Python id": 4338, "finished": true, "Ev Idx": 7238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905407.324, "dur": 5.536,
    "args": {
      "Python parent id": 4334, "Python id": 4339, "finished": true, "Ev Idx": 7239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905407.537, "dur": 5.272,
    "args": {
      "Python parent id": 4339, "Python id": 4340, "finished": true, "Ev Idx": 7240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905407.700, "dur": 3.305,
    "args": {
      "Python parent id": 4340, "Python id": 4341, "finished": true, "Ev Idx": 7241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905407.854, "dur": 3.087,
    "args": {
      "Python parent id": 4341, "Python id": 4342, "finished": true, "Ev Idx": 7242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905407.970, "dur": 0.274,
    "args": {
      "Python parent id": 4342, "Python id": 4343, "finished": true, "Ev Idx": 7243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905408.099, "dur": 0.061,
    "args": {
      "Python parent id": 4343, "Python id": 4344, "finished": true, "Ev Idx": 7244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905408.445, "dur": 2.128,
    "args": {
      "Python parent id": 4342, "Python id": 4345, "finished": true, "Ev Idx": 7245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905408.642, "dur": 1.848,
    "args": {
      "Python parent id": 4345, "Python id": 4346, "finished": true, "Ev Idx": 7246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905408.837, "dur": 1.577,
    "args": {
      "Python parent id": 4346, "Python id": 4347, "finished": true, "Ev Idx": 7247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905408.994, "dur": 1.173,
    "args": {
      "Python parent id": 4347, "Python id": 4348, "finished": true, "Ev Idx": 7248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905409.166, "dur": 0.529,
    "args": {
      "Python parent id": 4348, "Python id": 4349, "finished": true, "Ev Idx": 7249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905409.244, "dur": 0.038,
    "args": {
      "Python parent id": 4349, "Python id": 4350, "finished": true, "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905409.442, "dur": 0.184,
    "args": {
      "Python parent id": 4349, "Python id": 4351, "finished": true, "Ev Idx": 7251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905410.750, "dur": 0.145,
    "args": {
      "Python parent id": 4342, "Python id": 4352, "finished": true, "Ev Idx": 7252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905411.196, "dur": 0.102,
    "args": {
      "Python parent id": 4340, "Python id": 4353, "finished": true, "Ev Idx": 7253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905411.591, "dur": 1.159,
    "args": {
      "Python parent id": 4340, "Python id": 4354, "finished": true, "Ev Idx": 7254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905411.747, "dur": 0.966,
    "args": {
      "Python parent id": 4354, "Python id": 4355, "finished": true, "Ev Idx": 7255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905411.880, "dur": 0.366,
    "args": {
      "Python parent id": 4355, "Python id": 4356, "finished": true, "Ev Idx": 7256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905412.020, "dur": 0.167,
    "args": {
      "Python parent id": 4356, "Python id": 4357, "finished": true, "Ev Idx": 7257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905412.107, "dur": 0.049,
    "args": {
      "Python parent id": 4357, "Python id": 4358, "finished": true, "Ev Idx": 7258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905412.361, "dur": 0.292,
    "args": {
      "Python parent id": 4355, "Python id": 4359, "finished": true, "Ev Idx": 7259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905413.164, "dur": 0.175,
    "args": {
      "Python parent id": 4324, "Python id": 4360, "finished": true, "Ev Idx": 7260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905413.934, "dur": 1.109,
    "args": {
      "Python parent id": 4324, "Python id": 4361, "finished": true, "Ev Idx": 7261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905414.398, "dur": 0.578,
    "args": {
      "Python parent id": 4361, "Python id": 4362, "finished": true, "Ev Idx": 7262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905415.572, "dur": 3.426,
    "args": {
      "Python parent id": 4323, "Python id": 4363, "finished": true, "Ev Idx": 7263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905419.732, "dur": 8.499,
    "args": {
      "Python parent id": 12, "Python id": 4364, "finished": true, "Ev Idx": 7264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905428.673, "dur": 0.124,
    "args": {
      "Python parent id": 12, "Python id": 4365, "finished": true, "Ev Idx": 7265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905429.472, "dur": 0.163,
    "args": {
      "Python parent id": 12, "Python id": 4366, "finished": true, "Ev Idx": 7266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905430.112, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 4367, "finished": true, "Ev Idx": 7267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905430.534, "dur": 10.000,
    "args": {
      "Python parent id": 12, "Python id": 4368, "finished": true, "Ev Idx": 7268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905441.671, "dur": 13.879,
    "args": {
      "Python parent id": 12, "Python id": 4369, "finished": true, "Ev Idx": 7269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905441.984, "dur": 10.492,
    "args": {
      "Python parent id": 4369, "Python id": 4370, "finished": true, "Ev Idx": 7270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905442.119, "dur": 0.422,
    "args": {
      "Python parent id": 4370, "Python id": 4371, "finished": true, "Ev Idx": 7271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905442.262, "dur": 0.219,
    "args": {
      "Python parent id": 4371, "Python id": 4372, "finished": true, "Ev Idx": 7272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905442.354, "dur": 0.093,
    "args": {
      "Python parent id": 4372, "Python id": 4373, "finished": true, "Ev Idx": 7273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905442.839, "dur": 7.925,
    "args": {
      "Python parent id": 4370, "Python id": 4374, "finished": true, "Ev Idx": 7274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905442.920, "dur": 0.066,
    "args": {
      "Python parent id": 4374, "Python id": 4375, "finished": true, "Ev Idx": 7275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905443.117, "dur": 0.053,
    "args": {
      "Python parent id": 4374, "Python id": 4376, "finished": true, "Ev Idx": 7276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905443.304, "dur": 0.053,
    "args": {
      "Python parent id": 4374, "Python id": 4377, "finished": true, "Ev Idx": 7277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905443.740, "dur": 0.067,
    "args": {
      "Python parent id": 4374, "Python id": 4378, "finished": true, "Ev Idx": 7278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905444.063, "dur": 0.069,
    "args": {
      "Python parent id": 4374, "Python id": 4379, "finished": true, "Ev Idx": 7279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905444.424, "dur": 6.268,
    "args": {
      "Python parent id": 4374, "Python id": 4380, "finished": true, "Ev Idx": 7280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905444.520, "dur": 0.060,
    "args": {
      "Python parent id": 4380, "Python id": 4381, "finished": true, "Ev Idx": 7281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905444.736, "dur": 0.045,
    "args": {
      "Python parent id": 4380, "Python id": 4382, "finished": true, "Ev Idx": 7282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905444.929, "dur": 0.060,
    "args": {
      "Python parent id": 4380, "Python id": 4383, "finished": true, "Ev Idx": 7283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905445.257, "dur": 0.035,
    "args": {
      "Python parent id": 4380, "Python id": 4384, "finished": true, "Ev Idx": 7284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905445.571, "dur": 4.997,
    "args": {
      "Python parent id": 4380, "Python id": 4385, "finished": true, "Ev Idx": 7285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905445.766, "dur": 4.749,
    "args": {
      "Python parent id": 4385, "Python id": 4386, "finished": true, "Ev Idx": 7286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905445.907, "dur": 3.014,
    "args": {
      "Python parent id": 4386, "Python id": 4387, "finished": true, "Ev Idx": 7287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905446.115, "dur": 2.766,
    "args": {
      "Python parent id": 4387, "Python id": 4388, "finished": true, "Ev Idx": 7288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905446.252, "dur": 0.295,
    "args": {
      "Python parent id": 4388, "Python id": 4389, "finished": true, "Ev Idx": 7289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905446.389, "dur": 0.081,
    "args": {
      "Python parent id": 4389, "Python id": 4390, "finished": true, "Ev Idx": 7290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905446.698, "dur": 1.919,
    "args": {
      "Python parent id": 4388, "Python id": 4391, "finished": true, "Ev Idx": 7291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905446.860, "dur": 1.667,
    "args": {
      "Python parent id": 4391, "Python id": 4392, "finished": true, "Ev Idx": 7292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905447.022, "dur": 1.427,
    "args": {
      "Python parent id": 4392, "Python id": 4393, "finished": true, "Ev Idx": 7293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905447.139, "dur": 1.043,
    "args": {
      "Python parent id": 4393, "Python id": 4394, "finished": true, "Ev Idx": 7294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905447.280, "dur": 0.495,
    "args": {
      "Python parent id": 4394, "Python id": 4395, "finished": true, "Ev Idx": 7295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905447.352, "dur": 0.044,
    "args": {
      "Python parent id": 4395, "Python id": 4396, "finished": true, "Ev Idx": 7296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905447.583, "dur": 0.151,
    "args": {
      "Python parent id": 4395, "Python id": 4397, "finished": true, "Ev Idx": 7297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905448.784, "dur": 0.064,
    "args": {
      "Python parent id": 4388, "Python id": 4398, "finished": true, "Ev Idx": 7298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.062, "dur": 0.080,
    "args": {
      "Python parent id": 4386, "Python id": 4399, "finished": true, "Ev Idx": 7299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.428, "dur": 1.042,
    "args": {
      "Python parent id": 4386, "Python id": 4400, "finished": true, "Ev Idx": 7300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.551, "dur": 0.876,
    "args": {
      "Python parent id": 4400, "Python id": 4401, "finished": true, "Ev Idx": 7301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.690, "dur": 0.360,
    "args": {
      "Python parent id": 4401, "Python id": 4402, "finished": true, "Ev Idx": 7302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.824, "dur": 0.164,
    "args": {
      "Python parent id": 4402, "Python id": 4403, "finished": true, "Ev Idx": 7303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905449.909, "dur": 0.050,
    "args": {
      "Python parent id": 4403, "Python id": 4404, "finished": true, "Ev Idx": 7304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905450.179, "dur": 0.222,
    "args": {
      "Python parent id": 4401, "Python id": 4405, "finished": true, "Ev Idx": 7305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905450.875, "dur": 0.156,
    "args": {
      "Python parent id": 4370, "Python id": 4406, "finished": true, "Ev Idx": 7306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905451.537, "dur": 0.826,
    "args": {
      "Python parent id": 4370, "Python id": 4407, "finished": true, "Ev Idx": 7307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905451.936, "dur": 0.381,
    "args": {
      "Python parent id": 4407, "Python id": 4408, "finished": true, "Ev Idx": 7308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905452.762, "dur": 2.741,
    "args": {
      "Python parent id": 4369, "Python id": 4409, "finished": true, "Ev Idx": 7309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905456.504, "dur": 85.259,
    "args": {
      "Python parent id": 12, "Python id": 4410, "finished": true, "Ev Idx": 7310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905456.933, "dur": 0.436,
    "args": {
      "Python parent id": 4410, "Python id": 4411, "finished": true, "Ev Idx": 7311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905457.206, "dur": 0.117,
    "args": {
      "Python parent id": 4411, "Python id": 4412, "finished": true, "Ev Idx": 7312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905457.701, "dur": 0.351,
    "args": {
      "Python parent id": 4410, "Python id": 4413, "finished": true, "Ev Idx": 7313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905458.233, "dur": 0.285,
    "args": {
      "Python parent id": 4410, "Python id": 4414, "finished": true, "Ev Idx": 7314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905458.749, "dur": 1.558,
    "args": {
      "Python parent id": 4410, "Python id": 4415, "finished": true, "Ev Idx": 7315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905460.936, "dur": 80.705,
    "args": {
      "Python parent id": 4410, "Python id": 4416, "finished": true, "Ev Idx": 7316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905461.662, "dur": 79.800,
    "args": {
      "Python parent id": 4416, "Python id": 4417, "finished": true, "Ev Idx": 7317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905461.930, "dur": 79.456,
    "args": {
      "Python parent id": 4417, "Python id": 4418, "finished": true, "Ev Idx": 7318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905467.102, "dur": 73.083,
    "args": {
      "Python parent id": 4418, "Python id": 4419, "finished": true, "Ev Idx": 7319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905467.431, "dur": 72.673,
    "args": {
      "Python parent id": 4419, "Python id": 4420, "finished": true, "Ev Idx": 7320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905468.001, "dur": 71.968,
    "args": {
      "Python parent id": 4420, "Python id": 4421, "finished": true, "Ev Idx": 7321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905468.537, "dur": 0.427,
    "args": {
      "Python parent id": 4421, "Python id": 4422, "finished": true, "Ev Idx": 7322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905468.849, "dur": 0.057,
    "args": {
      "Python parent id": 4422, "Python id": 4423, "finished": true, "Ev Idx": 7323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905469.257, "dur": 0.340,
    "args": {
      "Python parent id": 4421, "Python id": 4424, "finished": true, "Ev Idx": 7324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905469.494, "dur": 0.057,
    "args": {
      "Python parent id": 4424, "Python id": 4425, "finished": true, "Ev Idx": 7325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905469.894, "dur": 69.958,
    "args": {
      "Python parent id": 4421, "Python id": 4426, "finished": true, "Ev Idx": 7326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905470.428, "dur": 68.287,
    "args": {
      "Python parent id": 4426, "Python id": 4427, "finished": true, "Ev Idx": 7327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905478.239, "dur": 16.862,
    "args": {
      "Python parent id": 4427, "Python id": 4428, "finished": true, "Ev Idx": 7328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905478.355, "dur": 0.108,
    "args": {
      "Python parent id": 4428, "Python id": 4429, "finished": true, "Ev Idx": 7329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905478.750, "dur": 16.259,
    "args": {
      "Python parent id": 4428, "Python id": 4430, "finished": true, "Ev Idx": 7330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905478.890, "dur": 7.216,
    "args": {
      "Python parent id": 4430, "Python id": 4431, "finished": true, "Ev Idx": 7331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905502.306, "dur": 7.550,
    "args": {
      "Python parent id": 4427, "Python id": 4432, "finished": true, "Ev Idx": 7332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905517.676, "dur": 7.513,
    "args": {
      "Python parent id": 4427, "Python id": 4433, "finished": true, "Ev Idx": 7333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905542.871, "dur": 15.541,
    "args": {
      "Python parent id": 12, "Python id": 4434, "finished": true, "Ev Idx": 7334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905543.196, "dur": 11.860,
    "args": {
      "Python parent id": 4434, "Python id": 4435, "finished": true, "Ev Idx": 7335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905543.396, "dur": 0.554,
    "args": {
      "Python parent id": 4435, "Python id": 4436, "finished": true, "Ev Idx": 7336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905543.621, "dur": 0.250,
    "args": {
      "Python parent id": 4436, "Python id": 4437, "finished": true, "Ev Idx": 7337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905543.724, "dur": 0.112,
    "args": {
      "Python parent id": 4437, "Python id": 4438, "finished": true, "Ev Idx": 7338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905544.318, "dur": 8.648,
    "args": {
      "Python parent id": 4435, "Python id": 4439, "finished": true, "Ev Idx": 7339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905544.432, "dur": 0.070,
    "args": {
      "Python parent id": 4439, "Python id": 4440, "finished": true, "Ev Idx": 7340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905544.627, "dur": 0.075,
    "args": {
      "Python parent id": 4439, "Python id": 4441, "finished": true, "Ev Idx": 7341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905544.834, "dur": 0.054,
    "args": {
      "Python parent id": 4439, "Python id": 4442, "finished": true, "Ev Idx": 7342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905545.323, "dur": 0.034,
    "args": {
      "Python parent id": 4439, "Python id": 4443, "finished": true, "Ev Idx": 7343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905545.629, "dur": 0.087,
    "args": {
      "Python parent id": 4439, "Python id": 4444, "finished": true, "Ev Idx": 7344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905546.047, "dur": 6.854,
    "args": {
      "Python parent id": 4439, "Python id": 4445, "finished": true, "Ev Idx": 7345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905546.133, "dur": 0.062,
    "args": {
      "Python parent id": 4445, "Python id": 4446, "finished": true, "Ev Idx": 7346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905546.336, "dur": 0.054,
    "args": {
      "Python parent id": 4445, "Python id": 4447, "finished": true, "Ev Idx": 7347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905546.541, "dur": 0.060,
    "args": {
      "Python parent id": 4445, "Python id": 4448, "finished": true, "Ev Idx": 7348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905546.931, "dur": 0.034,
    "args": {
      "Python parent id": 4445, "Python id": 4449, "finished": true, "Ev Idx": 7349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905547.292, "dur": 5.500,
    "args": {
      "Python parent id": 4445, "Python id": 4450, "finished": true, "Ev Idx": 7350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905547.577, "dur": 5.136,
    "args": {
      "Python parent id": 4450, "Python id": 4451, "finished": true, "Ev Idx": 7351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905547.730, "dur": 3.285,
    "args": {
      "Python parent id": 4451, "Python id": 4452, "finished": true, "Ev Idx": 7352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905547.937, "dur": 3.036,
    "args": {
      "Python parent id": 4452, "Python id": 4453, "finished": true, "Ev Idx": 7353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905548.077, "dur": 0.304,
    "args": {
      "Python parent id": 4453, "Python id": 4454, "finished": true, "Ev Idx": 7354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905548.240, "dur": 0.069,
    "args": {
      "Python parent id": 4454, "Python id": 4455, "finished": true, "Ev Idx": 7355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905548.541, "dur": 2.102,
    "args": {
      "Python parent id": 4453, "Python id": 4456, "finished": true, "Ev Idx": 7356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905548.756, "dur": 1.822,
    "args": {
      "Python parent id": 4456, "Python id": 4457, "finished": true, "Ev Idx": 7357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905548.957, "dur": 1.561,
    "args": {
      "Python parent id": 4457, "Python id": 4458, "finished": true, "Ev Idx": 7358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905549.098, "dur": 1.218,
    "args": {
      "Python parent id": 4458, "Python id": 4459, "finished": true, "Ev Idx": 7359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905549.290, "dur": 0.571,
    "args": {
      "Python parent id": 4459, "Python id": 4460, "finished": true, "Ev Idx": 7360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905549.380, "dur": 0.050,
    "args": {
      "Python parent id": 4460, "Python id": 4461, "finished": true, "Ev Idx": 7361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905549.611, "dur": 0.202,
    "args": {
      "Python parent id": 4460, "Python id": 4462, "finished": true, "Ev Idx": 7362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905550.809, "dur": 0.132,
    "args": {
      "Python parent id": 4453, "Python id": 4463, "finished": true, "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905551.186, "dur": 0.069,
    "args": {
      "Python parent id": 4451, "Python id": 4464, "finished": true, "Ev Idx": 7364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905551.566, "dur": 1.104,
    "args": {
      "Python parent id": 4451, "Python id": 4465, "finished": true, "Ev Idx": 7365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905551.698, "dur": 0.930,
    "args": {
      "Python parent id": 4465, "Python id": 4466, "finished": true, "Ev Idx": 7366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905551.835, "dur": 0.358,
    "args": {
      "Python parent id": 4466, "Python id": 4467, "finished": true, "Ev Idx": 7367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905551.976, "dur": 0.178,
    "args": {
      "Python parent id": 4467, "Python id": 4468, "finished": true, "Ev Idx": 7368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905552.062, "dur": 0.053,
    "args": {
      "Python parent id": 4468, "Python id": 4469, "finished": true, "Ev Idx": 7369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905552.312, "dur": 0.284,
    "args": {
      "Python parent id": 4466, "Python id": 4470, "finished": true, "Ev Idx": 7370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905553.081, "dur": 0.195,
    "args": {
      "Python parent id": 4435, "Python id": 4471, "finished": true, "Ev Idx": 7371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905553.868, "dur": 1.035,
    "args": {
      "Python parent id": 4435, "Python id": 4472, "finished": true, "Ev Idx": 7372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905554.347, "dur": 0.507,
    "args": {
      "Python parent id": 4472, "Python id": 4473, "finished": true, "Ev Idx": 7373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905555.408, "dur": 2.952,
    "args": {
      "Python parent id": 4434, "Python id": 4474, "finished": true, "Ev Idx": 7374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905558.977, "dur": 7.496,
    "args": {
      "Python parent id": 12, "Python id": 4475, "finished": true, "Ev Idx": 7375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905566.915, "dur": 0.089,
    "args": {
      "Python parent id": 12, "Python id": 4476, "finished": true, "Ev Idx": 7376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905567.622, "dur": 0.182,
    "args": {
      "Python parent id": 12, "Python id": 4477, "finished": true, "Ev Idx": 7377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905568.375, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 4478, "finished": true, "Ev Idx": 7378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905568.796, "dur": 8.671,
    "args": {
      "Python parent id": 12, "Python id": 4479, "finished": true, "Ev Idx": 7379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905578.589, "dur": 13.759,
    "args": {
      "Python parent id": 12, "Python id": 4480, "finished": true, "Ev Idx": 7380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905578.925, "dur": 10.312,
    "args": {
      "Python parent id": 4480, "Python id": 4481, "finished": true, "Ev Idx": 7381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905579.129, "dur": 0.460,
    "args": {
      "Python parent id": 4481, "Python id": 4482, "finished": true, "Ev Idx": 7382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905579.269, "dur": 0.230,
    "args": {
      "Python parent id": 4482, "Python id": 4483, "finished": true, "Ev Idx": 7383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905579.360, "dur": 0.103,
    "args": {
      "Python parent id": 4483, "Python id": 4484, "finished": true, "Ev Idx": 7384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905579.846, "dur": 7.649,
    "args": {
      "Python parent id": 4481, "Python id": 4485, "finished": true, "Ev Idx": 7385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905579.940, "dur": 0.064,
    "args": {
      "Python parent id": 4485, "Python id": 4486, "finished": true, "Ev Idx": 7386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905580.125, "dur": 0.050,
    "args": {
      "Python parent id": 4485, "Python id": 4487, "finished": true, "Ev Idx": 7387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905580.304, "dur": 0.051,
    "args": {
      "Python parent id": 4485, "Python id": 4488, "finished": true, "Ev Idx": 7388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905580.718, "dur": 0.039,
    "args": {
      "Python parent id": 4485, "Python id": 4489, "finished": true, "Ev Idx": 7389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905581.031, "dur": 0.067,
    "args": {
      "Python parent id": 4485, "Python id": 4490, "finished": true, "Ev Idx": 7390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905581.448, "dur": 6.000,
    "args": {
      "Python parent id": 4485, "Python id": 4491, "finished": true, "Ev Idx": 7391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905581.567, "dur": 0.059,
    "args": {
      "Python parent id": 4491, "Python id": 4492, "finished": true, "Ev Idx": 7392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905581.759, "dur": 0.054,
    "args": {
      "Python parent id": 4491, "Python id": 4493, "finished": true, "Ev Idx": 7393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905581.969, "dur": 0.059,
    "args": {
      "Python parent id": 4491, "Python id": 4494, "finished": true, "Ev Idx": 7394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905582.312, "dur": 0.036,
    "args": {
      "Python parent id": 4491, "Python id": 4495, "finished": true, "Ev Idx": 7395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905582.668, "dur": 4.667,
    "args": {
      "Python parent id": 4491, "Python id": 4496, "finished": true, "Ev Idx": 7396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905582.911, "dur": 4.371,
    "args": {
      "Python parent id": 4496, "Python id": 4497, "finished": true, "Ev Idx": 7397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.021, "dur": 2.689,
    "args": {
      "Python parent id": 4497, "Python id": 4498, "finished": true, "Ev Idx": 7398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.177, "dur": 2.489,
    "args": {
      "Python parent id": 4498, "Python id": 4499, "finished": true, "Ev Idx": 7399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.298, "dur": 0.229,
    "args": {
      "Python parent id": 4499, "Python id": 4500, "finished": true, "Ev Idx": 7400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.409, "dur": 0.070,
    "args": {
      "Python parent id": 4500, "Python id": 4501, "finished": true, "Ev Idx": 7401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.682, "dur": 1.716,
    "args": {
      "Python parent id": 4499, "Python id": 4502, "finished": true, "Ev Idx": 7402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905583.869, "dur": 1.474,
    "args": {
      "Python parent id": 4502, "Python id": 4503, "finished": true, "Ev Idx": 7403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905584.018, "dur": 1.263,
    "args": {
      "Python parent id": 4503, "Python id": 4504, "finished": true, "Ev Idx": 7404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905584.154, "dur": 0.969,
    "args": {
      "Python parent id": 4504, "Python id": 4505, "finished": true, "Ev Idx": 7405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905584.292, "dur": 0.443,
    "args": {
      "Python parent id": 4505, "Python id": 4506, "finished": true, "Ev Idx": 7406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905584.381, "dur": 0.041,
    "args": {
      "Python parent id": 4506, "Python id": 4507, "finished": true, "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905584.576, "dur": 0.122,
    "args": {
      "Python parent id": 4506, "Python id": 4508, "finished": true, "Ev Idx": 7408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905585.567, "dur": 0.064,
    "args": {
      "Python parent id": 4499, "Python id": 4509, "finished": true, "Ev Idx": 7409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905585.848, "dur": 0.104,
    "args": {
      "Python parent id": 4497, "Python id": 4510, "finished": true, "Ev Idx": 7410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.215, "dur": 1.023,
    "args": {
      "Python parent id": 4497, "Python id": 4511, "finished": true, "Ev Idx": 7411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.336, "dur": 0.859,
    "args": {
      "Python parent id": 4511, "Python id": 4512, "finished": true, "Ev Idx": 7412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.472, "dur": 0.357,
    "args": {
      "Python parent id": 4512, "Python id": 4513, "finished": true, "Ev Idx": 7413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.618, "dur": 0.168,
    "args": {
      "Python parent id": 4513, "Python id": 4514, "finished": true, "Ev Idx": 7414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.706, "dur": 0.049,
    "args": {
      "Python parent id": 4514, "Python id": 4515, "finished": true, "Ev Idx": 7415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905586.945, "dur": 0.207,
    "args": {
      "Python parent id": 4512, "Python id": 4516, "finished": true, "Ev Idx": 7416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905587.609, "dur": 0.164,
    "args": {
      "Python parent id": 4481, "Python id": 4517, "finished": true, "Ev Idx": 7417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905588.301, "dur": 0.803,
    "args": {
      "Python parent id": 4481, "Python id": 4518, "finished": true, "Ev Idx": 7418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905588.664, "dur": 0.391,
    "args": {
      "Python parent id": 4518, "Python id": 4519, "finished": true, "Ev Idx": 7419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905589.484, "dur": 2.819,
    "args": {
      "Python parent id": 4480, "Python id": 4520, "finished": true, "Ev Idx": 7420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905593.247, "dur": 80.107,
    "args": {
      "Python parent id": 12, "Python id": 4521, "finished": true, "Ev Idx": 7421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905593.619, "dur": 0.402,
    "args": {
      "Python parent id": 4521, "Python id": 4522, "finished": true, "Ev Idx": 7422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905593.880, "dur": 0.096,
    "args": {
      "Python parent id": 4522, "Python id": 4523, "finished": true, "Ev Idx": 7423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905594.300, "dur": 0.427,
    "args": {
      "Python parent id": 4521, "Python id": 4524, "finished": true, "Ev Idx": 7424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905594.936, "dur": 0.238,
    "args": {
      "Python parent id": 4521, "Python id": 4525, "finished": true, "Ev Idx": 7425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905595.431, "dur": 1.596,
    "args": {
      "Python parent id": 4521, "Python id": 4526, "finished": true, "Ev Idx": 7426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905597.580, "dur": 75.669,
    "args": {
      "Python parent id": 4521, "Python id": 4527, "finished": true, "Ev Idx": 7427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905598.408, "dur": 74.670,
    "args": {
      "Python parent id": 4527, "Python id": 4528, "finished": true, "Ev Idx": 7428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905598.672, "dur": 74.356,
    "args": {
      "Python parent id": 4528, "Python id": 4529, "finished": true, "Ev Idx": 7429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905603.552, "dur": 68.285,
    "args": {
      "Python parent id": 4529, "Python id": 4530, "finished": true, "Ev Idx": 7430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905603.804, "dur": 67.973,
    "args": {
      "Python parent id": 4530, "Python id": 4531, "finished": true, "Ev Idx": 7431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905604.344, "dur": 67.340,
    "args": {
      "Python parent id": 4531, "Python id": 4532, "finished": true, "Ev Idx": 7432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905604.871, "dur": 0.346,
    "args": {
      "Python parent id": 4532, "Python id": 4533, "finished": true, "Ev Idx": 7433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905605.104, "dur": 0.059,
    "args": {
      "Python parent id": 4533, "Python id": 4534, "finished": true, "Ev Idx": 7434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905605.472, "dur": 0.323,
    "args": {
      "Python parent id": 4532, "Python id": 4535, "finished": true, "Ev Idx": 7435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905605.702, "dur": 0.043,
    "args": {
      "Python parent id": 4535, "Python id": 4536, "finished": true, "Ev Idx": 7436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905606.071, "dur": 65.519,
    "args": {
      "Python parent id": 4532, "Python id": 4537, "finished": true, "Ev Idx": 7437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905606.598, "dur": 63.852,
    "args": {
      "Python parent id": 4537, "Python id": 4538, "finished": true, "Ev Idx": 7438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905612.981, "dur": 15.824,
    "args": {
      "Python parent id": 4538, "Python id": 4539, "finished": true, "Ev Idx": 7439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905613.103, "dur": 0.080,
    "args": {
      "Python parent id": 4539, "Python id": 4540, "finished": true, "Ev Idx": 7440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905613.467, "dur": 15.276,
    "args": {
      "Python parent id": 4539, "Python id": 4541, "finished": true, "Ev Idx": 7441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905613.602, "dur": 6.705,
    "args": {
      "Python parent id": 4541, "Python id": 4542, "finished": true, "Ev Idx": 7442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905635.534, "dur": 6.062,
    "args": {
      "Python parent id": 4538, "Python id": 4543, "finished": true, "Ev Idx": 7443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905649.171, "dur": 8.360,
    "args": {
      "Python parent id": 4538, "Python id": 4544, "finished": true, "Ev Idx": 7444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905674.573, "dur": 15.947,
    "args": {
      "Python parent id": 12, "Python id": 4545, "finished": true, "Ev Idx": 7445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905674.866, "dur": 12.222,
    "args": {
      "Python parent id": 4545, "Python id": 4546, "finished": true, "Ev Idx": 7446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905675.070, "dur": 0.515,
    "args": {
      "Python parent id": 4546, "Python id": 4547, "finished": true, "Ev Idx": 7447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905675.283, "dur": 0.242,
    "args": {
      "Python parent id": 4547, "Python id": 4548, "finished": true, "Ev Idx": 7448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905675.395, "dur": 0.077,
    "args": {
      "Python parent id": 4548, "Python id": 4549, "finished": true, "Ev Idx": 7449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905675.889, "dur": 9.203,
    "args": {
      "Python parent id": 4546, "Python id": 4550, "finished": true, "Ev Idx": 7450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905675.984, "dur": 0.099,
    "args": {
      "Python parent id": 4550, "Python id": 4551, "finished": true, "Ev Idx": 7451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905676.204, "dur": 0.075,
    "args": {
      "Python parent id": 4550, "Python id": 4552, "finished": true, "Ev Idx": 7452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905676.439, "dur": 0.052,
    "args": {
      "Python parent id": 4550, "Python id": 4553, "finished": true, "Ev Idx": 7453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905676.954, "dur": 0.078,
    "args": {
      "Python parent id": 4550, "Python id": 4554, "finished": true, "Ev Idx": 7454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905677.317, "dur": 0.090,
    "args": {
      "Python parent id": 4550, "Python id": 4555, "finished": true, "Ev Idx": 7455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905677.787, "dur": 7.257,
    "args": {
      "Python parent id": 4550, "Python id": 4556, "finished": true, "Ev Idx": 7456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905677.883, "dur": 0.059,
    "args": {
      "Python parent id": 4556, "Python id": 4557, "finished": true, "Ev Idx": 7457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905678.101, "dur": 0.054,
    "args": {
      "Python parent id": 4556, "Python id": 4558, "finished": true, "Ev Idx": 7458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905678.372, "dur": 0.066,
    "args": {
      "Python parent id": 4556, "Python id": 4559, "finished": true, "Ev Idx": 7459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905678.811, "dur": 0.030,
    "args": {
      "Python parent id": 4556, "Python id": 4560, "finished": true, "Ev Idx": 7460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905679.167, "dur": 5.796,
    "args": {
      "Python parent id": 4556, "Python id": 4561, "finished": true, "Ev Idx": 7461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905679.492, "dur": 5.390,
    "args": {
      "Python parent id": 4561, "Python id": 4562, "finished": true, "Ev Idx": 7462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905679.614, "dur": 3.541,
    "args": {
      "Python parent id": 4562, "Python id": 4563, "finished": true, "Ev Idx": 7463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905679.806, "dur": 3.280,
    "args": {
      "Python parent id": 4563, "Python id": 4564, "finished": true, "Ev Idx": 7464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905679.955, "dur": 0.298,
    "args": {
      "Python parent id": 4564, "Python id": 4565, "finished": true, "Ev Idx": 7465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905680.084, "dur": 0.082,
    "args": {
      "Python parent id": 4565, "Python id": 4566, "finished": true, "Ev Idx": 7466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905680.443, "dur": 2.222,
    "args": {
      "Python parent id": 4564, "Python id": 4567, "finished": true, "Ev Idx": 7467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905680.663, "dur": 1.918,
    "args": {
      "Python parent id": 4567, "Python id": 4568, "finished": true, "Ev Idx": 7468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905680.861, "dur": 1.639,
    "args": {
      "Python parent id": 4568, "Python id": 4569, "finished": true, "Ev Idx": 7469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905681.056, "dur": 1.212,
    "args": {
      "Python parent id": 4569, "Python id": 4570, "finished": true, "Ev Idx": 7470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905681.251, "dur": 0.514,
    "args": {
      "Python parent id": 4570, "Python id": 4571, "finished": true, "Ev Idx": 7471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905681.340, "dur": 0.043,
    "args": {
      "Python parent id": 4571, "Python id": 4572, "finished": true, "Ev Idx": 7472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905681.594, "dur": 0.132,
    "args": {
      "Python parent id": 4571, "Python id": 4573, "finished": true, "Ev Idx": 7473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905682.859, "dur": 0.155,
    "args": {
      "Python parent id": 4564, "Python id": 4574, "finished": true, "Ev Idx": 7474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905683.346, "dur": 0.070,
    "args": {
      "Python parent id": 4562, "Python id": 4575, "finished": true, "Ev Idx": 7475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905683.703, "dur": 1.120,
    "args": {
      "Python parent id": 4562, "Python id": 4576, "finished": true, "Ev Idx": 7476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905683.842, "dur": 0.944,
    "args": {
      "Python parent id": 4576, "Python id": 4577, "finished": true, "Ev Idx": 7477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905683.962, "dur": 0.377,
    "args": {
      "Python parent id": 4577, "Python id": 4578, "finished": true, "Ev Idx": 7478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905684.100, "dur": 0.167,
    "args": {
      "Python parent id": 4578, "Python id": 4579, "finished": true, "Ev Idx": 7479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905684.184, "dur": 0.049,
    "args": {
      "Python parent id": 4579, "Python id": 4580, "finished": true, "Ev Idx": 7480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905684.470, "dur": 0.257,
    "args": {
      "Python parent id": 4577, "Python id": 4581, "finished": true, "Ev Idx": 7481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905685.207, "dur": 0.184,
    "args": {
      "Python parent id": 4546, "Python id": 4582, "finished": true, "Ev Idx": 7482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905685.968, "dur": 0.971,
    "args": {
      "Python parent id": 4546, "Python id": 4583, "finished": true, "Ev Idx": 7483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905686.452, "dur": 0.440,
    "args": {
      "Python parent id": 4583, "Python id": 4584, "finished": true, "Ev Idx": 7484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905687.399, "dur": 3.080,
    "args": {
      "Python parent id": 4545, "Python id": 4585, "finished": true, "Ev Idx": 7485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905691.115, "dur": 6.993,
    "args": {
      "Python parent id": 12, "Python id": 4586, "finished": true, "Ev Idx": 7486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905698.591, "dur": 0.106,
    "args": {
      "Python parent id": 12, "Python id": 4587, "finished": true, "Ev Idx": 7487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905699.274, "dur": 0.189,
    "args": {
      "Python parent id": 12, "Python id": 4588, "finished": true, "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905699.926, "dur": 0.064,
    "args": {
      "Python parent id": 12, "Python id": 4589, "finished": true, "Ev Idx": 7489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905700.332, "dur": 8.805,
    "args": {
      "Python parent id": 12, "Python id": 4590, "finished": true, "Ev Idx": 7490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905710.382, "dur": 15.358,
    "args": {
      "Python parent id": 12, "Python id": 4591, "finished": true, "Ev Idx": 7491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905710.673, "dur": 12.070,
    "args": {
      "Python parent id": 4591, "Python id": 4592, "finished": true, "Ev Idx": 7492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905710.812, "dur": 0.452,
    "args": {
      "Python parent id": 4592, "Python id": 4593, "finished": true, "Ev Idx": 7493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905710.951, "dur": 0.241,
    "args": {
      "Python parent id": 4593, "Python id": 4594, "finished": true, "Ev Idx": 7494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905711.043, "dur": 0.116,
    "args": {
      "Python parent id": 4594, "Python id": 4595, "finished": true, "Ev Idx": 7495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905711.522, "dur": 9.435,
    "args": {
      "Python parent id": 4592, "Python id": 4596, "finished": true, "Ev Idx": 7496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905711.599, "dur": 0.060,
    "args": {
      "Python parent id": 4596, "Python id": 4597, "finished": true, "Ev Idx": 7497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905711.766, "dur": 0.076,
    "args": {
      "Python parent id": 4596, "Python id": 4598, "finished": true, "Ev Idx": 7498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905711.978, "dur": 0.049,
    "args": {
      "Python parent id": 4596, "Python id": 4599, "finished": true, "Ev Idx": 7499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905712.457, "dur": 0.035,
    "args": {
      "Python parent id": 4596, "Python id": 4600, "finished": true, "Ev Idx": 7500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905712.768, "dur": 0.068,
    "args": {
      "Python parent id": 4596, "Python id": 4601, "finished": true, "Ev Idx": 7501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905713.164, "dur": 7.738,
    "args": {
      "Python parent id": 4596, "Python id": 4602, "finished": true, "Ev Idx": 7502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905713.249, "dur": 0.058,
    "args": {
      "Python parent id": 4602, "Python id": 4603, "finished": true, "Ev Idx": 7503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905713.454, "dur": 0.052,
    "args": {
      "Python parent id": 4602, "Python id": 4604, "finished": true, "Ev Idx": 7504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905713.669, "dur": 0.065,
    "args": {
      "Python parent id": 4602, "Python id": 4605, "finished": true, "Ev Idx": 7505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905714.038, "dur": 0.032,
    "args": {
      "Python parent id": 4602, "Python id": 4606, "finished": true, "Ev Idx": 7506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905714.405, "dur": 6.413,
    "args": {
      "Python parent id": 4602, "Python id": 4607, "finished": true, "Ev Idx": 7507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905714.617, "dur": 6.102,
    "args": {
      "Python parent id": 4607, "Python id": 4608, "finished": true, "Ev Idx": 7508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905714.813, "dur": 2.989,
    "args": {
      "Python parent id": 4608, "Python id": 4609, "finished": true, "Ev Idx": 7509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905715.032, "dur": 2.727,
    "args": {
      "Python parent id": 4609, "Python id": 4610, "finished": true, "Ev Idx": 7510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905715.175, "dur": 0.249,
    "args": {
      "Python parent id": 4610, "Python id": 4611, "finished": true, "Ev Idx": 7511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905715.277, "dur": 0.067,
    "args": {
      "Python parent id": 4611, "Python id": 4612, "finished": true, "Ev Idx": 7512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905715.647, "dur": 1.777,
    "args": {
      "Python parent id": 4610, "Python id": 4613, "finished": true, "Ev Idx": 7513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905715.801, "dur": 1.568,
    "args": {
      "Python parent id": 4613, "Python id": 4614, "finished": true, "Ev Idx": 7514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905716.003, "dur": 1.319,
    "args": {
      "Python parent id": 4614, "Python id": 4615, "finished": true, "Ev Idx": 7515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905716.123, "dur": 1.034,
    "args": {
      "Python parent id": 4615, "Python id": 4616, "finished": true, "Ev Idx": 7516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905716.277, "dur": 0.431,
    "args": {
      "Python parent id": 4616, "Python id": 4617, "finished": true, "Ev Idx": 7517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905716.358, "dur": 0.044,
    "args": {
      "Python parent id": 4617, "Python id": 4618, "finished": true, "Ev Idx": 7518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905716.554, "dur": 0.112,
    "args": {
      "Python parent id": 4617, "Python id": 4619, "finished": true, "Ev Idx": 7519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905717.601, "dur": 0.112,
    "args": {
      "Python parent id": 4610, "Python id": 4620, "finished": true, "Ev Idx": 7520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905717.967, "dur": 0.082,
    "args": {
      "Python parent id": 4608, "Python id": 4621, "finished": true, "Ev Idx": 7521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905718.333, "dur": 2.341,
    "args": {
      "Python parent id": 4608, "Python id": 4622, "finished": true, "Ev Idx": 7522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905718.460, "dur": 2.171,
    "args": {
      "Python parent id": 4622, "Python id": 4623, "finished": true, "Ev Idx": 7523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905718.588, "dur": 1.504,
    "args": {
      "Python parent id": 4623, "Python id": 4624, "finished": true, "Ev Idx": 7524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905718.715, "dur": 1.316,
    "args": {
      "Python parent id": 4624, "Python id": 4625, "finished": true, "Ev Idx": 7525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905718.793, "dur": 1.215,
    "args": {
      "Python parent id": 4625, "Python id": 4626, "finished": true, "Ev Idx": 7526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905720.331, "dur": 0.255,
    "args": {
      "Python parent id": 4623, "Python id": 4627, "finished": true, "Ev Idx": 7527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905721.080, "dur": 0.175,
    "args": {
      "Python parent id": 4592, "Python id": 4628, "finished": true, "Ev Idx": 7528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905721.788, "dur": 0.832,
    "args": {
      "Python parent id": 4592, "Python id": 4629, "finished": true, "Ev Idx": 7529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905722.137, "dur": 0.437,
    "args": {
      "Python parent id": 4629, "Python id": 4630, "finished": true, "Ev Idx": 7530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905722.997, "dur": 2.697,
    "args": {
      "Python parent id": 4591, "Python id": 4631, "finished": true, "Ev Idx": 7531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905726.596, "dur": 79.908,
    "args": {
      "Python parent id": 12, "Python id": 4632, "finished": true, "Ev Idx": 7532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905726.958, "dur": 0.447,
    "args": {
      "Python parent id": 4632, "Python id": 4633, "finished": true, "Ev Idx": 7533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905727.247, "dur": 0.096,
    "args": {
      "Python parent id": 4633, "Python id": 4634, "finished": true, "Ev Idx": 7534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905727.723, "dur": 0.347,
    "args": {
      "Python parent id": 4632, "Python id": 4635, "finished": true, "Ev Idx": 7535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905728.234, "dur": 0.246,
    "args": {
      "Python parent id": 4632, "Python id": 4636, "finished": true, "Ev Idx": 7536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905728.684, "dur": 1.670,
    "args": {
      "Python parent id": 4632, "Python id": 4637, "finished": true, "Ev Idx": 7537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905730.869, "dur": 75.571,
    "args": {
      "Python parent id": 4632, "Python id": 4638, "finished": true, "Ev Idx": 7538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905731.766, "dur": 74.480,
    "args": {
      "Python parent id": 4638, "Python id": 4639, "finished": true, "Ev Idx": 7539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905732.046, "dur": 74.106,
    "args": {
      "Python parent id": 4639, "Python id": 4640, "finished": true, "Ev Idx": 7540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905737.044, "dur": 67.887,
    "args": {
      "Python parent id": 4640, "Python id": 4641, "finished": true, "Ev Idx": 7541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905737.286, "dur": 67.591,
    "args": {
      "Python parent id": 4641, "Python id": 4642, "finished": true, "Ev Idx": 7542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905737.849, "dur": 66.860,
    "args": {
      "Python parent id": 4642, "Python id": 4643, "finished": true, "Ev Idx": 7543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905738.282, "dur": 0.349,
    "args": {
      "Python parent id": 4643, "Python id": 4644, "finished": true, "Ev Idx": 7544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905738.508, "dur": 0.080,
    "args": {
      "Python parent id": 4644, "Python id": 4645, "finished": true, "Ev Idx": 7545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905738.968, "dur": 0.301,
    "args": {
      "Python parent id": 4643, "Python id": 4646, "finished": true, "Ev Idx": 7546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905739.176, "dur": 0.046,
    "args": {
      "Python parent id": 4646, "Python id": 4647, "finished": true, "Ev Idx": 7547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905739.577, "dur": 65.062,
    "args": {
      "Python parent id": 4643, "Python id": 4648, "finished": true, "Ev Idx": 7548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905740.250, "dur": 63.180,
    "args": {
      "Python parent id": 4648, "Python id": 4649, "finished": true, "Ev Idx": 7549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905746.792, "dur": 16.023,
    "args": {
      "Python parent id": 4649, "Python id": 4650, "finished": true, "Ev Idx": 7550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905746.933, "dur": 0.101,
    "args": {
      "Python parent id": 4650, "Python id": 4651, "finished": true, "Ev Idx": 7551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905747.328, "dur": 15.341,
    "args": {
      "Python parent id": 4650, "Python id": 4652, "finished": true, "Ev Idx": 7552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905747.466, "dur": 6.824,
    "args": {
      "Python parent id": 4652, "Python id": 4653, "finished": true, "Ev Idx": 7553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905769.920, "dur": 6.298,
    "args": {
      "Python parent id": 4649, "Python id": 4654, "finished": true, "Ev Idx": 7554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905783.595, "dur": 7.288,
    "args": {
      "Python parent id": 4649, "Python id": 4655, "finished": true, "Ev Idx": 7555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905807.625, "dur": 15.473,
    "args": {
      "Python parent id": 12, "Python id": 4656, "finished": true, "Ev Idx": 7556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905807.956, "dur": 11.333,
    "args": {
      "Python parent id": 4656, "Python id": 4657, "finished": true, "Ev Idx": 7557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905808.133, "dur": 0.465,
    "args": {
      "Python parent id": 4657, "Python id": 4658, "finished": true, "Ev Idx": 7558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905808.307, "dur": 0.204,
    "args": {
      "Python parent id": 4658, "Python id": 4659, "finished": true, "Ev Idx": 7559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905808.422, "dur": 0.060,
    "args": {
      "Python parent id": 4659, "Python id": 4660, "finished": true, "Ev Idx": 7560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905808.840, "dur": 8.449,
    "args": {
      "Python parent id": 4657, "Python id": 4661, "finished": true, "Ev Idx": 7561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905808.929, "dur": 0.074,
    "args": {
      "Python parent id": 4661, "Python id": 4662, "finished": true, "Ev Idx": 7562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905809.112, "dur": 0.056,
    "args": {
      "Python parent id": 4661, "Python id": 4663, "finished": true, "Ev Idx": 7563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905809.291, "dur": 0.056,
    "args": {
      "Python parent id": 4661, "Python id": 4664, "finished": true, "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905809.775, "dur": 0.028,
    "args": {
      "Python parent id": 4661, "Python id": 4665, "finished": true, "Ev Idx": 7565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905810.115, "dur": 0.081,
    "args": {
      "Python parent id": 4661, "Python id": 4666, "finished": true, "Ev Idx": 7566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905810.514, "dur": 6.721,
    "args": {
      "Python parent id": 4661, "Python id": 4667, "finished": true, "Ev Idx": 7567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905810.629, "dur": 0.065,
    "args": {
      "Python parent id": 4667, "Python id": 4668, "finished": true, "Ev Idx": 7568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905810.828, "dur": 0.054,
    "args": {
      "Python parent id": 4667, "Python id": 4669, "finished": true, "Ev Idx": 7569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905811.026, "dur": 0.066,
    "args": {
      "Python parent id": 4667, "Python id": 4670, "finished": true, "Ev Idx": 7570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905811.425, "dur": 0.066,
    "args": {
      "Python parent id": 4667, "Python id": 4671, "finished": true, "Ev Idx": 7571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905811.852, "dur": 5.274,
    "args": {
      "Python parent id": 4667, "Python id": 4672, "finished": true, "Ev Idx": 7572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905812.067, "dur": 5.004,
    "args": {
      "Python parent id": 4672, "Python id": 4673, "finished": true, "Ev Idx": 7573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905812.211, "dur": 3.203,
    "args": {
      "Python parent id": 4673, "Python id": 4674, "finished": true, "Ev Idx": 7574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905812.423, "dur": 2.932,
    "args": {
      "Python parent id": 4674, "Python id": 4675, "finished": true, "Ev Idx": 7575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905812.578, "dur": 0.284,
    "args": {
      "Python parent id": 4675, "Python id": 4676, "finished": true, "Ev Idx": 7576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905812.717, "dur": 0.069,
    "args": {
      "Python parent id": 4676, "Python id": 4677, "finished": true, "Ev Idx": 7577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.046, "dur": 1.935,
    "args": {
      "Python parent id": 4675, "Python id": 4678, "finished": true, "Ev Idx": 7578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.239, "dur": 1.680,
    "args": {
      "Python parent id": 4678, "Python id": 4679, "finished": true, "Ev Idx": 7579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.410, "dur": 1.443,
    "args": {
      "Python parent id": 4679, "Python id": 4680, "finished": true, "Ev Idx": 7580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.545, "dur": 1.121,
    "args": {
      "Python parent id": 4680, "Python id": 4681, "finished": true, "Ev Idx": 7581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.755, "dur": 0.446,
    "args": {
      "Python parent id": 4681, "Python id": 4682, "finished": true, "Ev Idx": 7582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905813.834, "dur": 0.045,
    "args": {
      "Python parent id": 4682, "Python id": 4683, "finished": true, "Ev Idx": 7583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905814.014, "dur": 0.146,
    "args": {
      "Python parent id": 4682, "Python id": 4684, "finished": true, "Ev Idx": 7584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905815.168, "dur": 0.153,
    "args": {
      "Python parent id": 4675, "Python id": 4685, "finished": true, "Ev Idx": 7585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905815.566, "dur": 0.096,
    "args": {
      "Python parent id": 4673, "Python id": 4686, "finished": true, "Ev Idx": 7586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905815.927, "dur": 1.101,
    "args": {
      "Python parent id": 4673, "Python id": 4687, "finished": true, "Ev Idx": 7587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905816.051, "dur": 0.933,
    "args": {
      "Python parent id": 4687, "Python id": 4688, "finished": true, "Ev Idx": 7588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905816.186, "dur": 0.354,
    "args": {
      "Python parent id": 4688, "Python id": 4689, "finished": true, "Ev Idx": 7589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905816.315, "dur": 0.171,
    "args": {
      "Python parent id": 4689, "Python id": 4690, "finished": true, "Ev Idx": 7590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905816.400, "dur": 0.053,
    "args": {
      "Python parent id": 4690, "Python id": 4691, "finished": true, "Ev Idx": 7591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905816.667, "dur": 0.283,
    "args": {
      "Python parent id": 4688, "Python id": 4692, "finished": true, "Ev Idx": 7592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905817.408, "dur": 0.218,
    "args": {
      "Python parent id": 4657, "Python id": 4693, "finished": true, "Ev Idx": 7593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905818.180, "dur": 0.968,
    "args": {
      "Python parent id": 4657, "Python id": 4694, "finished": true, "Ev Idx": 7594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905818.653, "dur": 0.424,
    "args": {
      "Python parent id": 4694, "Python id": 4695, "finished": true, "Ev Idx": 7595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905819.562, "dur": 3.490,
    "args": {
      "Python parent id": 4656, "Python id": 4696, "finished": true, "Ev Idx": 7596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905823.618, "dur": 7.466,
    "args": {
      "Python parent id": 12, "Python id": 4697, "finished": true, "Ev Idx": 7597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905831.526, "dur": 0.093,
    "args": {
      "Python parent id": 12, "Python id": 4698, "finished": true, "Ev Idx": 7598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905832.261, "dur": 0.197,
    "args": {
      "Python parent id": 12, "Python id": 4699, "finished": true, "Ev Idx": 7599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905832.965, "dur": 0.065,
    "args": {
      "Python parent id": 12, "Python id": 4700, "finished": true, "Ev Idx": 7600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905833.371, "dur": 10.133,
    "args": {
      "Python parent id": 12, "Python id": 4701, "finished": true, "Ev Idx": 7601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905844.610, "dur": 13.537,
    "args": {
      "Python parent id": 12, "Python id": 4702, "finished": true, "Ev Idx": 7602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905844.915, "dur": 10.327,
    "args": {
      "Python parent id": 4702, "Python id": 4703, "finished": true, "Ev Idx": 7603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905845.050, "dur": 0.445,
    "args": {
      "Python parent id": 4703, "Python id": 4704, "finished": true, "Ev Idx": 7604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905845.211, "dur": 0.214,
    "args": {
      "Python parent id": 4704, "Python id": 4705, "finished": true, "Ev Idx": 7605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905845.301, "dur": 0.091,
    "args": {
      "Python parent id": 4705, "Python id": 4706, "finished": true, "Ev Idx": 7606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905845.763, "dur": 7.803,
    "args": {
      "Python parent id": 4703, "Python id": 4707, "finished": true, "Ev Idx": 7607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905845.848, "dur": 0.064,
    "args": {
      "Python parent id": 4707, "Python id": 4708, "finished": true, "Ev Idx": 7608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905846.022, "dur": 0.052,
    "args": {
      "Python parent id": 4707, "Python id": 4709, "finished": true, "Ev Idx": 7609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905846.208, "dur": 0.052,
    "args": {
      "Python parent id": 4707, "Python id": 4710, "finished": true, "Ev Idx": 7610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905846.597, "dur": 0.089,
    "args": {
      "Python parent id": 4707, "Python id": 4711, "finished": true, "Ev Idx": 7611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905846.920, "dur": 0.089,
    "args": {
      "Python parent id": 4707, "Python id": 4712, "finished": true, "Ev Idx": 7612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905847.356, "dur": 6.160,
    "args": {
      "Python parent id": 4707, "Python id": 4713, "finished": true, "Ev Idx": 7613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905847.446, "dur": 0.063,
    "args": {
      "Python parent id": 4713, "Python id": 4714, "finished": true, "Ev Idx": 7614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905847.650, "dur": 0.053,
    "args": {
      "Python parent id": 4713, "Python id": 4715, "finished": true, "Ev Idx": 7615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905847.881, "dur": 0.069,
    "args": {
      "Python parent id": 4713, "Python id": 4716, "finished": true, "Ev Idx": 7616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905848.255, "dur": 0.065,
    "args": {
      "Python parent id": 4713, "Python id": 4717, "finished": true, "Ev Idx": 7617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905848.608, "dur": 4.818,
    "args": {
      "Python parent id": 4713, "Python id": 4718, "finished": true, "Ev Idx": 7618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905848.868, "dur": 4.506,
    "args": {
      "Python parent id": 4718, "Python id": 4719, "finished": true, "Ev Idx": 7619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905848.996, "dur": 2.817,
    "args": {
      "Python parent id": 4719, "Python id": 4720, "finished": true, "Ev Idx": 7620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905849.195, "dur": 2.553,
    "args": {
      "Python parent id": 4720, "Python id": 4721, "finished": true, "Ev Idx": 7621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905849.321, "dur": 0.259,
    "args": {
      "Python parent id": 4721, "Python id": 4722, "finished": true, "Ev Idx": 7622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905849.432, "dur": 0.062,
    "args": {
      "Python parent id": 4722, "Python id": 4723, "finished": true, "Ev Idx": 7623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905849.725, "dur": 1.758,
    "args": {
      "Python parent id": 4721, "Python id": 4724, "finished": true, "Ev Idx": 7624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905849.916, "dur": 1.511,
    "args": {
      "Python parent id": 4724, "Python id": 4725, "finished": true, "Ev Idx": 7625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905850.108, "dur": 1.255,
    "args": {
      "Python parent id": 4725, "Python id": 4726, "finished": true, "Ev Idx": 7626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905850.222, "dur": 0.947,
    "args": {
      "Python parent id": 4726, "Python id": 4727, "finished": true, "Ev Idx": 7627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905850.369, "dur": 0.415,
    "args": {
      "Python parent id": 4727, "Python id": 4728, "finished": true, "Ev Idx": 7628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905850.449, "dur": 0.040,
    "args": {
      "Python parent id": 4728, "Python id": 4729, "finished": true, "Ev Idx": 7629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905850.611, "dur": 0.133,
    "args": {
      "Python parent id": 4728, "Python id": 4730, "finished": true, "Ev Idx": 7630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905851.647, "dur": 0.070,
    "args": {
      "Python parent id": 4721, "Python id": 4731, "finished": true, "Ev Idx": 7631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905851.960, "dur": 0.112,
    "args": {
      "Python parent id": 4719, "Python id": 4732, "finished": true, "Ev Idx": 7632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905852.286, "dur": 1.044,
    "args": {
      "Python parent id": 4719, "Python id": 4733, "finished": true, "Ev Idx": 7633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905852.444, "dur": 0.846,
    "args": {
      "Python parent id": 4733, "Python id": 4734, "finished": true, "Ev Idx": 7634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905852.570, "dur": 0.353,
    "args": {
      "Python parent id": 4734, "Python id": 4735, "finished": true, "Ev Idx": 7635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905852.697, "dur": 0.167,
    "args": {
      "Python parent id": 4735, "Python id": 4736, "finished": true, "Ev Idx": 7636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905852.782, "dur": 0.050,
    "args": {
      "Python parent id": 4736, "Python id": 4737, "finished": true, "Ev Idx": 7637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905853.039, "dur": 0.222,
    "args": {
      "Python parent id": 4734, "Python id": 4738, "finished": true, "Ev Idx": 7638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905853.672, "dur": 0.120,
    "args": {
      "Python parent id": 4703, "Python id": 4739, "finished": true, "Ev Idx": 7639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905854.380, "dur": 0.751,
    "args": {
      "Python parent id": 4703, "Python id": 4740, "finished": true, "Ev Idx": 7640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905854.719, "dur": 0.362,
    "args": {
      "Python parent id": 4740, "Python id": 4741, "finished": true, "Ev Idx": 7641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905855.531, "dur": 2.570,
    "args": {
      "Python parent id": 4702, "Python id": 4742, "finished": true, "Ev Idx": 7642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905859.045, "dur": 82.088,
    "args": {
      "Python parent id": 12, "Python id": 4743, "finished": true, "Ev Idx": 7643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905859.339, "dur": 0.352,
    "args": {
      "Python parent id": 4743, "Python id": 4744, "finished": true, "Ev Idx": 7644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905859.564, "dur": 0.080,
    "args": {
      "Python parent id": 4744, "Python id": 4745, "finished": true, "Ev Idx": 7645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905860.173, "dur": 0.418,
    "args": {
      "Python parent id": 4743, "Python id": 4746, "finished": true, "Ev Idx": 7646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905860.798, "dur": 0.254,
    "args": {
      "Python parent id": 4743, "Python id": 4747, "finished": true, "Ev Idx": 7647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905861.256, "dur": 1.537,
    "args": {
      "Python parent id": 4743, "Python id": 4748, "finished": true, "Ev Idx": 7648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905863.327, "dur": 77.670,
    "args": {
      "Python parent id": 4743, "Python id": 4749, "finished": true, "Ev Idx": 7649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905864.149, "dur": 76.691,
    "args": {
      "Python parent id": 4749, "Python id": 4750, "finished": true, "Ev Idx": 7650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905864.397, "dur": 76.355,
    "args": {
      "Python parent id": 4750, "Python id": 4751, "finished": true, "Ev Idx": 7651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740905869.481, "dur": 70.105,
    "args": {
      "Python parent id": 4751, "Python id": 4752, "finished": true, "Ev Idx": 7652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740905869.792, "dur": 69.720,
    "args": {
      "Python parent id": 4752, "Python id": 4753, "finished": true, "Ev Idx": 7653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740905870.350, "dur": 68.967,
    "args": {
      "Python parent id": 4753, "Python id": 4754, "finished": true, "Ev Idx": 7654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905870.801, "dur": 0.415,
    "args": {
      "Python parent id": 4754, "Python id": 4755, "finished": true, "Ev Idx": 7655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905871.068, "dur": 0.086,
    "args": {
      "Python parent id": 4755, "Python id": 4756, "finished": true, "Ev Idx": 7656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905871.471, "dur": 0.285,
    "args": {
      "Python parent id": 4754, "Python id": 4757, "finished": true, "Ev Idx": 7657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905871.660, "dur": 0.051,
    "args": {
      "Python parent id": 4757, "Python id": 4758, "finished": true, "Ev Idx": 7658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905872.062, "dur": 67.177,
    "args": {
      "Python parent id": 4754, "Python id": 4759, "finished": true, "Ev Idx": 7659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740905872.655, "dur": 65.347,
    "args": {
      "Python parent id": 4759, "Python id": 4760, "finished": true, "Ev Idx": 7660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740905879.522, "dur": 16.793,
    "args": {
      "Python parent id": 4760, "Python id": 4761, "finished": true, "Ev Idx": 7661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740905879.672, "dur": 0.146,
    "args": {
      "Python parent id": 4761, "Python id": 4762, "finished": true, "Ev Idx": 7662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740905880.161, "dur": 16.034,
    "args": {
      "Python parent id": 4761, "Python id": 4763, "finished": true, "Ev Idx": 7663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905880.286, "dur": 7.156,
    "args": {
      "Python parent id": 4763, "Python id": 4764, "finished": true, "Ev Idx": 7664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905903.123, "dur": 6.191,
    "args": {
      "Python parent id": 4760, "Python id": 4765, "finished": true, "Ev Idx": 7665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740905917.616, "dur": 7.487,
    "args": {
      "Python parent id": 4760, "Python id": 4766, "finished": true, "Ev Idx": 7666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905942.333, "dur": 15.597,
    "args": {
      "Python parent id": 12, "Python id": 4767, "finished": true, "Ev Idx": 7667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905942.640, "dur": 11.617,
    "args": {
      "Python parent id": 4767, "Python id": 4768, "finished": true, "Ev Idx": 7668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905942.837, "dur": 0.452,
    "args": {
      "Python parent id": 4768, "Python id": 4769, "finished": true, "Ev Idx": 7669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905942.996, "dur": 0.233,
    "args": {
      "Python parent id": 4769, "Python id": 4770, "finished": true, "Ev Idx": 7670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905943.127, "dur": 0.069,
    "args": {
      "Python parent id": 4770, "Python id": 4771, "finished": true, "Ev Idx": 7671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905943.514, "dur": 8.652,
    "args": {
      "Python parent id": 4768, "Python id": 4772, "finished": true, "Ev Idx": 7672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905943.609, "dur": 0.066,
    "args": {
      "Python parent id": 4772, "Python id": 4773, "finished": true, "Ev Idx": 7673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905943.799, "dur": 0.069,
    "args": {
      "Python parent id": 4772, "Python id": 4774, "finished": true, "Ev Idx": 7674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905944.003, "dur": 0.056,
    "args": {
      "Python parent id": 4772, "Python id": 4775, "finished": true, "Ev Idx": 7675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905944.520, "dur": 0.025,
    "args": {
      "Python parent id": 4772, "Python id": 4776, "finished": true, "Ev Idx": 7676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905944.800, "dur": 0.085,
    "args": {
      "Python parent id": 4772, "Python id": 4777, "finished": true, "Ev Idx": 7677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905945.238, "dur": 6.856,
    "args": {
      "Python parent id": 4772, "Python id": 4778, "finished": true, "Ev Idx": 7678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905945.339, "dur": 0.058,
    "args": {
      "Python parent id": 4778, "Python id": 4779, "finished": true, "Ev Idx": 7679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905945.546, "dur": 0.056,
    "args": {
      "Python parent id": 4778, "Python id": 4780, "finished": true, "Ev Idx": 7680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905945.807, "dur": 0.064,
    "args": {
      "Python parent id": 4778, "Python id": 4781, "finished": true, "Ev Idx": 7681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905946.242, "dur": 0.050,
    "args": {
      "Python parent id": 4778, "Python id": 4782, "finished": true, "Ev Idx": 7682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905946.607, "dur": 5.371,
    "args": {
      "Python parent id": 4778, "Python id": 4783, "finished": true, "Ev Idx": 7683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905946.844, "dur": 5.083,
    "args": {
      "Python parent id": 4783, "Python id": 4784, "finished": true, "Ev Idx": 7684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905946.993, "dur": 3.199,
    "args": {
      "Python parent id": 4784, "Python id": 4785, "finished": true, "Ev Idx": 7685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905947.172, "dur": 2.976,
    "args": {
      "Python parent id": 4785, "Python id": 4786, "finished": true, "Ev Idx": 7686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905947.311, "dur": 0.277,
    "args": {
      "Python parent id": 4786, "Python id": 4787, "finished": true, "Ev Idx": 7687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905947.425, "dur": 0.117,
    "args": {
      "Python parent id": 4787, "Python id": 4788, "finished": true, "Ev Idx": 7688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905947.772, "dur": 2.000,
    "args": {
      "Python parent id": 4786, "Python id": 4789, "finished": true, "Ev Idx": 7689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905947.972, "dur": 1.741,
    "args": {
      "Python parent id": 4789, "Python id": 4790, "finished": true, "Ev Idx": 7690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905948.129, "dur": 1.502,
    "args": {
      "Python parent id": 4790, "Python id": 4791, "finished": true, "Ev Idx": 7691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905948.268, "dur": 1.138,
    "args": {
      "Python parent id": 4791, "Python id": 4792, "finished": true, "Ev Idx": 7692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905948.462, "dur": 0.479,
    "args": {
      "Python parent id": 4792, "Python id": 4793, "finished": true, "Ev Idx": 7693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905948.558, "dur": 0.049,
    "args": {
      "Python parent id": 4793, "Python id": 4794, "finished": true, "Ev Idx": 7694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905948.753, "dur": 0.146,
    "args": {
      "Python parent id": 4793, "Python id": 4795, "finished": true, "Ev Idx": 7695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905949.973, "dur": 0.139,
    "args": {
      "Python parent id": 4786, "Python id": 4796, "finished": true, "Ev Idx": 7696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905950.393, "dur": 0.100,
    "args": {
      "Python parent id": 4784, "Python id": 4797, "finished": true, "Ev Idx": 7697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905950.741, "dur": 1.124,
    "args": {
      "Python parent id": 4784, "Python id": 4798, "finished": true, "Ev Idx": 7698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905950.930, "dur": 0.894,
    "args": {
      "Python parent id": 4798, "Python id": 4799, "finished": true, "Ev Idx": 7699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905951.059, "dur": 0.357,
    "args": {
      "Python parent id": 4799, "Python id": 4800, "finished": true, "Ev Idx": 7700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905951.192, "dur": 0.170,
    "args": {
      "Python parent id": 4800, "Python id": 4801, "finished": true, "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905951.272, "dur": 0.057,
    "args": {
      "Python parent id": 4801, "Python id": 4802, "finished": true, "Ev Idx": 7702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905951.534, "dur": 0.235,
    "args": {
      "Python parent id": 4799, "Python id": 4803, "finished": true, "Ev Idx": 7703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905952.290, "dur": 0.200,
    "args": {
      "Python parent id": 4768, "Python id": 4804, "finished": true, "Ev Idx": 7704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905953.064, "dur": 1.040,
    "args": {
      "Python parent id": 4768, "Python id": 4805, "finished": true, "Ev Idx": 7705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905953.549, "dur": 0.506,
    "args": {
      "Python parent id": 4805, "Python id": 4806, "finished": true, "Ev Idx": 7706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905954.587, "dur": 3.299,
    "args": {
      "Python parent id": 4767, "Python id": 4807, "finished": true, "Ev Idx": 7707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905958.626, "dur": 8.157,
    "args": {
      "Python parent id": 12, "Python id": 4808, "finished": true, "Ev Idx": 7708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740905967.249, "dur": 0.133,
    "args": {
      "Python parent id": 12, "Python id": 4809, "finished": true, "Ev Idx": 7709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905968.028, "dur": 0.211,
    "args": {
      "Python parent id": 12, "Python id": 4810, "finished": true, "Ev Idx": 7710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905968.791, "dur": 0.065,
    "args": {
      "Python parent id": 12, "Python id": 4811, "finished": true, "Ev Idx": 7711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740905969.318, "dur": 9.068,
    "args": {
      "Python parent id": 12, "Python id": 4812, "finished": true, "Ev Idx": 7712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740905979.587, "dur": 14.049,
    "args": {
      "Python parent id": 12, "Python id": 4813, "finished": true, "Ev Idx": 7713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740905979.901, "dur": 10.860,
    "args": {
      "Python parent id": 4813, "Python id": 4814, "finished": true, "Ev Idx": 7714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905980.061, "dur": 0.444,
    "args": {
      "Python parent id": 4814, "Python id": 4815, "finished": true, "Ev Idx": 7715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905980.227, "dur": 0.207,
    "args": {
      "Python parent id": 4815, "Python id": 4816, "finished": true, "Ev Idx": 7716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905980.309, "dur": 0.090,
    "args": {
      "Python parent id": 4816, "Python id": 4817, "finished": true, "Ev Idx": 7717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905980.759, "dur": 8.153,
    "args": {
      "Python parent id": 4814, "Python id": 4818, "finished": true, "Ev Idx": 7718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905980.881, "dur": 0.059,
    "args": {
      "Python parent id": 4818, "Python id": 4819, "finished": true, "Ev Idx": 7719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905981.093, "dur": 0.055,
    "args": {
      "Python parent id": 4818, "Python id": 4820, "finished": true, "Ev Idx": 7720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905981.321, "dur": 0.059,
    "args": {
      "Python parent id": 4818, "Python id": 4821, "finished": true, "Ev Idx": 7721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905981.796, "dur": 0.037,
    "args": {
      "Python parent id": 4818, "Python id": 4822, "finished": true, "Ev Idx": 7722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905982.107, "dur": 0.075,
    "args": {
      "Python parent id": 4818, "Python id": 4823, "finished": true, "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905982.510, "dur": 6.349,
    "args": {
      "Python parent id": 4818, "Python id": 4824, "finished": true, "Ev Idx": 7724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905982.610, "dur": 0.062,
    "args": {
      "Python parent id": 4824, "Python id": 4825, "finished": true, "Ev Idx": 7725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905982.819, "dur": 0.060,
    "args": {
      "Python parent id": 4824, "Python id": 4826, "finished": true, "Ev Idx": 7726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905983.064, "dur": 0.065,
    "args": {
      "Python parent id": 4824, "Python id": 4827, "finished": true, "Ev Idx": 7727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740905983.495, "dur": 0.033,
    "args": {
      "Python parent id": 4824, "Python id": 4828, "finished": true, "Ev Idx": 7728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740905983.917, "dur": 4.843,
    "args": {
      "Python parent id": 4824, "Python id": 4829, "finished": true, "Ev Idx": 7729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.080, "dur": 4.627,
    "args": {
      "Python parent id": 4829, "Python id": 4830, "finished": true, "Ev Idx": 7730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.209, "dur": 2.916,
    "args": {
      "Python parent id": 4830, "Python id": 4831, "finished": true, "Ev Idx": 7731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.389, "dur": 2.694,
    "args": {
      "Python parent id": 4831, "Python id": 4832, "finished": true, "Ev Idx": 7732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.519, "dur": 0.246,
    "args": {
      "Python parent id": 4832, "Python id": 4833, "finished": true, "Ev Idx": 7733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.632, "dur": 0.065,
    "args": {
      "Python parent id": 4833, "Python id": 4834, "finished": true, "Ev Idx": 7734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740905984.931, "dur": 1.854,
    "args": {
      "Python parent id": 4832, "Python id": 4835, "finished": true, "Ev Idx": 7735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.131, "dur": 1.599,
    "args": {
      "Python parent id": 4835, "Python id": 4836, "finished": true, "Ev Idx": 7736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.286, "dur": 1.381,
    "args": {
      "Python parent id": 4836, "Python id": 4837, "finished": true, "Ev Idx": 7737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.445, "dur": 1.003,
    "args": {
      "Python parent id": 4837, "Python id": 4838, "finished": true, "Ev Idx": 7738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.585, "dur": 0.466,
    "args": {
      "Python parent id": 4838, "Python id": 4839, "finished": true, "Ev Idx": 7739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.674, "dur": 0.046,
    "args": {
      "Python parent id": 4839, "Python id": 4840, "finished": true, "Ev Idx": 7740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740905985.850, "dur": 0.152,
    "args": {
      "Python parent id": 4839, "Python id": 4841, "finished": true, "Ev Idx": 7741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740905986.989, "dur": 0.061,
    "args": {
      "Python parent id": 4832, "Python id": 4842, "finished": true, "Ev Idx": 7742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740905987.292, "dur": 0.073,
    "args": {
      "Python parent id": 4830, "Python id": 4843, "finished": true, "Ev Idx": 7743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740905987.584, "dur": 1.083,
    "args": {
      "Python parent id": 4830, "Python id": 4844, "finished": true, "Ev Idx": 7744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740905987.739, "dur": 0.888,
    "args": {
      "Python parent id": 4844, "Python id": 4845, "finished": true, "Ev Idx": 7745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740905987.862, "dur": 0.356,
    "args": {
      "Python parent id": 4845, "Python id": 4846, "finished": true, "Ev Idx": 7746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740905987.988, "dur": 0.173,
    "args": {
      "Python parent id": 4846, "Python id": 4847, "finished": true, "Ev Idx": 7747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740905988.069, "dur": 0.063,
    "args": {
      "Python parent id": 4847, "Python id": 4848, "finished": true, "Ev Idx": 7748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740905988.346, "dur": 0.255,
    "args": {
      "Python parent id": 4845, "Python id": 4849, "finished": true, "Ev Idx": 7749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740905989.029, "dur": 0.168,
    "args": {
      "Python parent id": 4814, "Python id": 4850, "finished": true, "Ev Idx": 7750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740905989.730, "dur": 0.875,
    "args": {
      "Python parent id": 4814, "Python id": 4851, "finished": true, "Ev Idx": 7751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740905990.158, "dur": 0.379,
    "args": {
      "Python parent id": 4851, "Python id": 4852, "finished": true, "Ev Idx": 7752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740905990.991, "dur": 2.596,
    "args": {
      "Python parent id": 4813, "Python id": 4853, "finished": true, "Ev Idx": 7753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740905994.479, "dur": 94.872,
    "args": {
      "Python parent id": 12, "Python id": 4854, "finished": true, "Ev Idx": 7754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740905994.788, "dur": 0.334,
    "args": {
      "Python parent id": 4854, "Python id": 4855, "finished": true, "Ev Idx": 7755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740905995.008, "dur": 0.066,
    "args": {
      "Python parent id": 4855, "Python id": 4856, "finished": true, "Ev Idx": 7756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905995.398, "dur": 0.398,
    "args": {
      "Python parent id": 4854, "Python id": 4857, "finished": true, "Ev Idx": 7757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740905996.004, "dur": 0.286,
    "args": {
      "Python parent id": 4854, "Python id": 4858, "finished": true, "Ev Idx": 7758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740905996.500, "dur": 1.477,
    "args": {
      "Python parent id": 4854, "Python id": 4859, "finished": true, "Ev Idx": 7759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740905998.567, "dur": 90.692,
    "args": {
      "Python parent id": 4854, "Python id": 4860, "finished": true, "Ev Idx": 7760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740905999.330, "dur": 89.712,
    "args": {
      "Python parent id": 4860, "Python id": 4861, "finished": true, "Ev Idx": 7761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740905999.542, "dur": 89.399,
    "args": {
      "Python parent id": 4861, "Python id": 4862, "finished": true, "Ev Idx": 7762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906004.283, "dur": 83.029,
    "args": {
      "Python parent id": 4862, "Python id": 4863, "finished": true, "Ev Idx": 7763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906004.538, "dur": 82.730,
    "args": {
      "Python parent id": 4863, "Python id": 4864, "finished": true, "Ev Idx": 7764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906005.091, "dur": 82.013,
    "args": {
      "Python parent id": 4864, "Python id": 4865, "finished": true, "Ev Idx": 7765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906005.629, "dur": 0.397,
    "args": {
      "Python parent id": 4865, "Python id": 4866, "finished": true, "Ev Idx": 7766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906005.903, "dur": 0.081,
    "args": {
      "Python parent id": 4866, "Python id": 4867, "finished": true, "Ev Idx": 7767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906006.282, "dur": 0.305,
    "args": {
      "Python parent id": 4865, "Python id": 4868, "finished": true, "Ev Idx": 7768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906006.496, "dur": 0.048,
    "args": {
      "Python parent id": 4868, "Python id": 4869, "finished": true, "Ev Idx": 7769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906006.890, "dur": 80.128,
    "args": {
      "Python parent id": 4865, "Python id": 4870, "finished": true, "Ev Idx": 7770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906007.543, "dur": 77.817,
    "args": {
      "Python parent id": 4870, "Python id": 4871, "finished": true, "Ev Idx": 7771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906014.044, "dur": 15.553,
    "args": {
      "Python parent id": 4871, "Python id": 4872, "finished": true, "Ev Idx": 7772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906014.200, "dur": 0.127,
    "args": {
      "Python parent id": 4872, "Python id": 4873, "finished": true, "Ev Idx": 7773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906014.623, "dur": 14.839,
    "args": {
      "Python parent id": 4872, "Python id": 4874, "finished": true, "Ev Idx": 7774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906014.738, "dur": 6.685,
    "args": {
      "Python parent id": 4874, "Python id": 4875, "finished": true, "Ev Idx": 7775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906036.174, "dur": 6.354,
    "args": {
      "Python parent id": 4871, "Python id": 4876, "finished": true, "Ev Idx": 7776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906050.175, "dur": 7.392,
    "args": {
      "Python parent id": 4871, "Python id": 4877, "finished": true, "Ev Idx": 7777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906090.933, "dur": 18.518,
    "args": {
      "Python parent id": 12, "Python id": 4878, "finished": true, "Ev Idx": 7778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906091.325, "dur": 13.861,
    "args": {
      "Python parent id": 4878, "Python id": 4879, "finished": true, "Ev Idx": 7779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906091.516, "dur": 0.640,
    "args": {
      "Python parent id": 4879, "Python id": 4880, "finished": true, "Ev Idx": 7780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906091.723, "dur": 0.346,
    "args": {
      "Python parent id": 4880, "Python id": 4881, "finished": true, "Ev Idx": 7781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906091.879, "dur": 0.126,
    "args": {
      "Python parent id": 4881, "Python id": 4882, "finished": true, "Ev Idx": 7782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906092.573, "dur": 9.827,
    "args": {
      "Python parent id": 4879, "Python id": 4883, "finished": true, "Ev Idx": 7783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906092.704, "dur": 0.102,
    "args": {
      "Python parent id": 4883, "Python id": 4884, "finished": true, "Ev Idx": 7784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906092.898, "dur": 0.059,
    "args": {
      "Python parent id": 4883, "Python id": 4885, "finished": true, "Ev Idx": 7785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906093.120, "dur": 0.059,
    "args": {
      "Python parent id": 4883, "Python id": 4886, "finished": true, "Ev Idx": 7786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906093.588, "dur": 0.062,
    "args": {
      "Python parent id": 4883, "Python id": 4887, "finished": true, "Ev Idx": 7787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906093.970, "dur": 0.074,
    "args": {
      "Python parent id": 4883, "Python id": 4888, "finished": true, "Ev Idx": 7788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906094.399, "dur": 7.925,
    "args": {
      "Python parent id": 4883, "Python id": 4889, "finished": true, "Ev Idx": 7789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906094.525, "dur": 0.059,
    "args": {
      "Python parent id": 4889, "Python id": 4890, "finished": true, "Ev Idx": 7790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906094.769, "dur": 0.061,
    "args": {
      "Python parent id": 4889, "Python id": 4891, "finished": true, "Ev Idx": 7791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906095.007, "dur": 0.067,
    "args": {
      "Python parent id": 4889, "Python id": 4892, "finished": true, "Ev Idx": 7792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906095.453, "dur": 0.037,
    "args": {
      "Python parent id": 4889, "Python id": 4893, "finished": true, "Ev Idx": 7793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906095.822, "dur": 6.433,
    "args": {
      "Python parent id": 4889, "Python id": 4894, "finished": true, "Ev Idx": 7794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.078, "dur": 6.087,
    "args": {
      "Python parent id": 4894, "Python id": 4895, "finished": true, "Ev Idx": 7795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.196, "dur": 3.929,
    "args": {
      "Python parent id": 4895, "Python id": 4896, "finished": true, "Ev Idx": 7796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.396, "dur": 3.638,
    "args": {
      "Python parent id": 4896, "Python id": 4897, "finished": true, "Ev Idx": 7797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.520, "dur": 0.310,
    "args": {
      "Python parent id": 4897, "Python id": 4898, "finished": true, "Ev Idx": 7798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.663, "dur": 0.080,
    "args": {
      "Python parent id": 4898, "Python id": 4899, "finished": true, "Ev Idx": 7799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906096.981, "dur": 2.630,
    "args": {
      "Python parent id": 4897, "Python id": 4900, "finished": true, "Ev Idx": 7800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906097.198, "dur": 2.284,
    "args": {
      "Python parent id": 4900, "Python id": 4901, "finished": true, "Ev Idx": 7801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906097.401, "dur": 2.022,
    "args": {
      "Python parent id": 4901, "Python id": 4902, "finished": true, "Ev Idx": 7802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906097.574, "dur": 1.547,
    "args": {
      "Python parent id": 4902, "Python id": 4903, "finished": true, "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906097.793, "dur": 0.530,
    "args": {
      "Python parent id": 4903, "Python id": 4904, "finished": true, "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906097.873, "dur": 0.044,
    "args": {
      "Python parent id": 4904, "Python id": 4905, "finished": true, "Ev Idx": 7805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906098.094, "dur": 0.167,
    "args": {
      "Python parent id": 4904, "Python id": 4906, "finished": true, "Ev Idx": 7806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906099.793, "dur": 0.166,
    "args": {
      "Python parent id": 4897, "Python id": 4907, "finished": true, "Ev Idx": 7807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906100.325, "dur": 0.126,
    "args": {
      "Python parent id": 4895, "Python id": 4908, "finished": true, "Ev Idx": 7808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906100.785, "dur": 1.317,
    "args": {
      "Python parent id": 4895, "Python id": 4909, "finished": true, "Ev Idx": 7809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906100.933, "dur": 1.129,
    "args": {
      "Python parent id": 4909, "Python id": 4910, "finished": true, "Ev Idx": 7810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906101.071, "dur": 0.423,
    "args": {
      "Python parent id": 4910, "Python id": 4911, "finished": true, "Ev Idx": 7811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906101.215, "dur": 0.191,
    "args": {
      "Python parent id": 4911, "Python id": 4912, "finished": true, "Ev Idx": 7812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906101.323, "dur": 0.054,
    "args": {
      "Python parent id": 4912, "Python id": 4913, "finished": true, "Ev Idx": 7813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906101.632, "dur": 0.371,
    "args": {
      "Python parent id": 4910, "Python id": 4914, "finished": true, "Ev Idx": 7814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906102.532, "dur": 0.318,
    "args": {
      "Python parent id": 4879, "Python id": 4915, "finished": true, "Ev Idx": 7815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906103.642, "dur": 1.352,
    "args": {
      "Python parent id": 4879, "Python id": 4916, "finished": true, "Ev Idx": 7816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906104.238, "dur": 0.667,
    "args": {
      "Python parent id": 4916, "Python id": 4917, "finished": true, "Ev Idx": 7817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906105.605, "dur": 3.748,
    "args": {
      "Python parent id": 4878, "Python id": 4918, "finished": true, "Ev Idx": 7818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906110.017, "dur": 10.311,
    "args": {
      "Python parent id": 12, "Python id": 4919, "finished": true, "Ev Idx": 7819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906120.773, "dur": 0.115,
    "args": {
      "Python parent id": 12, "Python id": 4920, "finished": true, "Ev Idx": 7820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906121.540, "dur": 0.279,
    "args": {
      "Python parent id": 12, "Python id": 4921, "finished": true, "Ev Idx": 7821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906122.310, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 4922, "finished": true, "Ev Idx": 7822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906122.732, "dur": 11.331,
    "args": {
      "Python parent id": 12, "Python id": 4923, "finished": true, "Ev Idx": 7823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906135.253, "dur": 15.385,
    "args": {
      "Python parent id": 12, "Python id": 4924, "finished": true, "Ev Idx": 7824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906135.529, "dur": 11.787,
    "args": {
      "Python parent id": 4924, "Python id": 4925, "finished": true, "Ev Idx": 7825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906135.740, "dur": 0.467,
    "args": {
      "Python parent id": 4925, "Python id": 4926, "finished": true, "Ev Idx": 7826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906135.889, "dur": 0.245,
    "args": {
      "Python parent id": 4926, "Python id": 4927, "finished": true, "Ev Idx": 7827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906136.015, "dur": 0.086,
    "args": {
      "Python parent id": 4927, "Python id": 4928, "finished": true, "Ev Idx": 7828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906136.493, "dur": 8.859,
    "args": {
      "Python parent id": 4925, "Python id": 4929, "finished": true, "Ev Idx": 7829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906136.572, "dur": 0.065,
    "args": {
      "Python parent id": 4929, "Python id": 4930, "finished": true, "Ev Idx": 7830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906136.739, "dur": 0.058,
    "args": {
      "Python parent id": 4929, "Python id": 4931, "finished": true, "Ev Idx": 7831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906137.001, "dur": 0.071,
    "args": {
      "Python parent id": 4929, "Python id": 4932, "finished": true, "Ev Idx": 7832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906137.546, "dur": 0.040,
    "args": {
      "Python parent id": 4929, "Python id": 4933, "finished": true, "Ev Idx": 7833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906137.845, "dur": 0.067,
    "args": {
      "Python parent id": 4929, "Python id": 4934, "finished": true, "Ev Idx": 7834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906138.277, "dur": 7.021,
    "args": {
      "Python parent id": 4929, "Python id": 4935, "finished": true, "Ev Idx": 7835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906138.385, "dur": 0.059,
    "args": {
      "Python parent id": 4935, "Python id": 4936, "finished": true, "Ev Idx": 7836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906138.623, "dur": 0.057,
    "args": {
      "Python parent id": 4935, "Python id": 4937, "finished": true, "Ev Idx": 7837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906138.846, "dur": 0.067,
    "args": {
      "Python parent id": 4935, "Python id": 4938, "finished": true, "Ev Idx": 7838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906139.334, "dur": 0.070,
    "args": {
      "Python parent id": 4935, "Python id": 4939, "finished": true, "Ev Idx": 7839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906139.681, "dur": 5.502,
    "args": {
      "Python parent id": 4935, "Python id": 4940, "finished": true, "Ev Idx": 7840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906139.906, "dur": 5.222,
    "args": {
      "Python parent id": 4940, "Python id": 4941, "finished": true, "Ev Idx": 7841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906140.067, "dur": 3.223,
    "args": {
      "Python parent id": 4941, "Python id": 4942, "finished": true, "Ev Idx": 7842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906140.325, "dur": 2.901,
    "args": {
      "Python parent id": 4942, "Python id": 4943, "finished": true, "Ev Idx": 7843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906140.440, "dur": 0.272,
    "args": {
      "Python parent id": 4943, "Python id": 4944, "finished": true, "Ev Idx": 7844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906140.587, "dur": 0.068,
    "args": {
      "Python parent id": 4944, "Python id": 4945, "finished": true, "Ev Idx": 7845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906140.876, "dur": 2.021,
    "args": {
      "Python parent id": 4943, "Python id": 4946, "finished": true, "Ev Idx": 7846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.044, "dur": 1.770,
    "args": {
      "Python parent id": 4946, "Python id": 4947, "finished": true, "Ev Idx": 7847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.258, "dur": 1.495,
    "args": {
      "Python parent id": 4947, "Python id": 4948, "finished": true, "Ev Idx": 7848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.395, "dur": 1.136,
    "args": {
      "Python parent id": 4948, "Python id": 4949, "finished": true, "Ev Idx": 7849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.582, "dur": 0.479,
    "args": {
      "Python parent id": 4949, "Python id": 4950, "finished": true, "Ev Idx": 7850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.657, "dur": 0.042,
    "args": {
      "Python parent id": 4950, "Python id": 4951, "finished": true, "Ev Idx": 7851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906141.869, "dur": 0.117,
    "args": {
      "Python parent id": 4950, "Python id": 4952, "finished": true, "Ev Idx": 7852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906143.068, "dur": 0.101,
    "args": {
      "Python parent id": 4943, "Python id": 4953, "finished": true, "Ev Idx": 7853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906143.459, "dur": 0.103,
    "args": {
      "Python parent id": 4941, "Python id": 4954, "finished": true, "Ev Idx": 7854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906143.939, "dur": 1.144,
    "args": {
      "Python parent id": 4941, "Python id": 4955, "finished": true, "Ev Idx": 7855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906144.122, "dur": 0.920,
    "args": {
      "Python parent id": 4955, "Python id": 4956, "finished": true, "Ev Idx": 7856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906144.245, "dur": 0.397,
    "args": {
      "Python parent id": 4956, "Python id": 4957, "finished": true, "Ev Idx": 7857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906144.374, "dur": 0.190,
    "args": {
      "Python parent id": 4957, "Python id": 4958, "finished": true, "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906144.479, "dur": 0.052,
    "args": {
      "Python parent id": 4958, "Python id": 4959, "finished": true, "Ev Idx": 7859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906144.762, "dur": 0.252,
    "args": {
      "Python parent id": 4956, "Python id": 4960, "finished": true, "Ev Idx": 7860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906145.482, "dur": 0.184,
    "args": {
      "Python parent id": 4925, "Python id": 4961, "finished": true, "Ev Idx": 7861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906146.300, "dur": 0.887,
    "args": {
      "Python parent id": 4925, "Python id": 4962, "finished": true, "Ev Idx": 7862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906146.694, "dur": 0.431,
    "args": {
      "Python parent id": 4962, "Python id": 4963, "finished": true, "Ev Idx": 7863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906147.611, "dur": 2.931,
    "args": {
      "Python parent id": 4924, "Python id": 4964, "finished": true, "Ev Idx": 7864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906151.596, "dur": 95.351,
    "args": {
      "Python parent id": 12, "Python id": 4965, "finished": true, "Ev Idx": 7865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906151.919, "dur": 0.473,
    "args": {
      "Python parent id": 4965, "Python id": 4966, "finished": true, "Ev Idx": 7866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906152.187, "dur": 0.142,
    "args": {
      "Python parent id": 4966, "Python id": 4967, "finished": true, "Ev Idx": 7867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906152.698, "dur": 0.533,
    "args": {
      "Python parent id": 4965, "Python id": 4968, "finished": true, "Ev Idx": 7868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906153.428, "dur": 0.258,
    "args": {
      "Python parent id": 4965, "Python id": 4969, "finished": true, "Ev Idx": 7869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906153.902, "dur": 1.621,
    "args": {
      "Python parent id": 4965, "Python id": 4970, "finished": true, "Ev Idx": 7870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906156.124, "dur": 90.756,
    "args": {
      "Python parent id": 4965, "Python id": 4971, "finished": true, "Ev Idx": 7871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906156.995, "dur": 89.679,
    "args": {
      "Python parent id": 4971, "Python id": 4972, "finished": true, "Ev Idx": 7872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906157.309, "dur": 89.309,
    "args": {
      "Python parent id": 4972, "Python id": 4973, "finished": true, "Ev Idx": 7873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906163.435, "dur": 81.832,
    "args": {
      "Python parent id": 4973, "Python id": 4974, "finished": true, "Ev Idx": 7874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906163.689, "dur": 81.527,
    "args": {
      "Python parent id": 4974, "Python id": 4975, "finished": true, "Ev Idx": 7875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906164.345, "dur": 80.738,
    "args": {
      "Python parent id": 4975, "Python id": 4976, "finished": true, "Ev Idx": 7876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906164.942, "dur": 0.377,
    "args": {
      "Python parent id": 4976, "Python id": 4977, "finished": true, "Ev Idx": 7877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906165.204, "dur": 0.073,
    "args": {
      "Python parent id": 4977, "Python id": 4978, "finished": true, "Ev Idx": 7878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906165.603, "dur": 0.329,
    "args": {
      "Python parent id": 4976, "Python id": 4979, "finished": true, "Ev Idx": 7879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906165.833, "dur": 0.048,
    "args": {
      "Python parent id": 4979, "Python id": 4980, "finished": true, "Ev Idx": 7880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906166.276, "dur": 78.719,
    "args": {
      "Python parent id": 4976, "Python id": 4981, "finished": true, "Ev Idx": 7881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906166.954, "dur": 76.791,
    "args": {
      "Python parent id": 4981, "Python id": 4982, "finished": true, "Ev Idx": 7882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906178.167, "dur": 18.505,
    "args": {
      "Python parent id": 4982, "Python id": 4983, "finished": true, "Ev Idx": 7883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906178.315, "dur": 0.154,
    "args": {
      "Python parent id": 4983, "Python id": 4984, "finished": true, "Ev Idx": 7884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906178.815, "dur": 17.772,
    "args": {
      "Python parent id": 4983, "Python id": 4985, "finished": true, "Ev Idx": 7885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906178.947, "dur": 7.746,
    "args": {
      "Python parent id": 4985, "Python id": 4986, "finished": true, "Ev Idx": 7886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906203.780, "dur": 7.706,
    "args": {
      "Python parent id": 4982, "Python id": 4987, "finished": true, "Ev Idx": 7887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906219.844, "dur": 8.474,
    "args": {
      "Python parent id": 4982, "Python id": 4988, "finished": true, "Ev Idx": 7888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906248.208, "dur": 17.797,
    "args": {
      "Python parent id": 12, "Python id": 4989, "finished": true, "Ev Idx": 7889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906248.552, "dur": 13.795,
    "args": {
      "Python parent id": 4989, "Python id": 4990, "finished": true, "Ev Idx": 7890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906248.733, "dur": 0.456,
    "args": {
      "Python parent id": 4990, "Python id": 4991, "finished": true, "Ev Idx": 7891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906248.916, "dur": 0.211,
    "args": {
      "Python parent id": 4991, "Python id": 4992, "finished": true, "Ev Idx": 7892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906249.030, "dur": 0.064,
    "args": {
      "Python parent id": 4992, "Python id": 4993, "finished": true, "Ev Idx": 7893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906249.522, "dur": 10.432,
    "args": {
      "Python parent id": 4990, "Python id": 4994, "finished": true, "Ev Idx": 7894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906249.650, "dur": 0.060,
    "args": {
      "Python parent id": 4994, "Python id": 4995, "finished": true, "Ev Idx": 7895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906249.826, "dur": 0.058,
    "args": {
      "Python parent id": 4994, "Python id": 4996, "finished": true, "Ev Idx": 7896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906250.021, "dur": 0.073,
    "args": {
      "Python parent id": 4994, "Python id": 4997, "finished": true, "Ev Idx": 7897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906250.510, "dur": 0.029,
    "args": {
      "Python parent id": 4994, "Python id": 4998, "finished": true, "Ev Idx": 7898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906250.780, "dur": 0.068,
    "args": {
      "Python parent id": 4994, "Python id": 4999, "finished": true, "Ev Idx": 7899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906251.211, "dur": 8.673,
    "args": {
      "Python parent id": 4994, "Python id": 5000, "finished": true, "Ev Idx": 7900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906251.312, "dur": 0.059,
    "args": {
      "Python parent id": 5000, "Python id": 5001, "finished": true, "Ev Idx": 7901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906252.806, "dur": 0.051,
    "args": {
      "Python parent id": 5000, "Python id": 5002, "finished": true, "Ev Idx": 7902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906253.004, "dur": 0.066,
    "args": {
      "Python parent id": 5000, "Python id": 5003, "finished": true, "Ev Idx": 7903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906253.420, "dur": 0.031,
    "args": {
      "Python parent id": 5000, "Python id": 5004, "finished": true, "Ev Idx": 7904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906253.750, "dur": 6.024,
    "args": {
      "Python parent id": 5000, "Python id": 5005, "finished": true, "Ev Idx": 7905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906254.054, "dur": 5.650,
    "args": {
      "Python parent id": 5005, "Python id": 5006, "finished": true, "Ev Idx": 7906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906254.179, "dur": 3.525,
    "args": {
      "Python parent id": 5006, "Python id": 5007, "finished": true, "Ev Idx": 7907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906254.372, "dur": 3.269,
    "args": {
      "Python parent id": 5007, "Python id": 5008, "finished": true, "Ev Idx": 7908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906254.505, "dur": 0.309,
    "args": {
      "Python parent id": 5008, "Python id": 5009, "finished": true, "Ev Idx": 7909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906254.674, "dur": 0.051,
    "args": {
      "Python parent id": 5009, "Python id": 5010, "finished": true, "Ev Idx": 7910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.023, "dur": 2.277,
    "args": {
      "Python parent id": 5008, "Python id": 5011, "finished": true, "Ev Idx": 7911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.227, "dur": 1.977,
    "args": {
      "Python parent id": 5011, "Python id": 5012, "finished": true, "Ev Idx": 7912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.429, "dur": 1.695,
    "args": {
      "Python parent id": 5012, "Python id": 5013, "finished": true, "Ev Idx": 7913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.614, "dur": 1.262,
    "args": {
      "Python parent id": 5013, "Python id": 5014, "finished": true, "Ev Idx": 7914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.810, "dur": 0.503,
    "args": {
      "Python parent id": 5014, "Python id": 5015, "finished": true, "Ev Idx": 7915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906255.887, "dur": 0.047,
    "args": {
      "Python parent id": 5015, "Python id": 5016, "finished": true, "Ev Idx": 7916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906256.109, "dur": 0.142,
    "args": {
      "Python parent id": 5015, "Python id": 5017, "finished": true, "Ev Idx": 7917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906257.465, "dur": 0.113,
    "args": {
      "Python parent id": 5008, "Python id": 5018, "finished": true, "Ev Idx": 7918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906257.890, "dur": 0.106,
    "args": {
      "Python parent id": 5006, "Python id": 5019, "finished": true, "Ev Idx": 7919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906258.283, "dur": 1.358,
    "args": {
      "Python parent id": 5006, "Python id": 5020, "finished": true, "Ev Idx": 7920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906258.443, "dur": 1.140,
    "args": {
      "Python parent id": 5020, "Python id": 5021, "finished": true, "Ev Idx": 7921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906258.599, "dur": 0.441,
    "args": {
      "Python parent id": 5021, "Python id": 5022, "finished": true, "Ev Idx": 7922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906258.776, "dur": 0.169,
    "args": {
      "Python parent id": 5022, "Python id": 5023, "finished": true, "Ev Idx": 7923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906258.863, "dur": 0.054,
    "args": {
      "Python parent id": 5023, "Python id": 5024, "finished": true, "Ev Idx": 7924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906259.181, "dur": 0.344,
    "args": {
      "Python parent id": 5021, "Python id": 5025, "finished": true, "Ev Idx": 7925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906260.087, "dur": 0.266,
    "args": {
      "Python parent id": 4990, "Python id": 5026, "finished": true, "Ev Idx": 7926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906261.017, "dur": 1.159,
    "args": {
      "Python parent id": 4990, "Python id": 5027, "finished": true, "Ev Idx": 7927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906261.506, "dur": 0.599,
    "args": {
      "Python parent id": 5027, "Python id": 5028, "finished": true, "Ev Idx": 7928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906262.757, "dur": 3.167,
    "args": {
      "Python parent id": 4989, "Python id": 5029, "finished": true, "Ev Idx": 7929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906266.590, "dur": 7.726,
    "args": {
      "Python parent id": 12, "Python id": 5030, "finished": true, "Ev Idx": 7930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906274.751, "dur": 0.140,
    "args": {
      "Python parent id": 12, "Python id": 5031, "finished": true, "Ev Idx": 7931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906275.486, "dur": 0.198,
    "args": {
      "Python parent id": 12, "Python id": 5032, "finished": true, "Ev Idx": 7932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906276.201, "dur": 0.070,
    "args": {
      "Python parent id": 12, "Python id": 5033, "finished": true, "Ev Idx": 7933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906276.633, "dur": 9.803,
    "args": {
      "Python parent id": 12, "Python id": 5034, "finished": true, "Ev Idx": 7934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906287.632, "dur": 14.239,
    "args": {
      "Python parent id": 12, "Python id": 5035, "finished": true, "Ev Idx": 7935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906287.994, "dur": 11.033,
    "args": {
      "Python parent id": 5035, "Python id": 5036, "finished": true, "Ev Idx": 7936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906288.175, "dur": 0.465,
    "args": {
      "Python parent id": 5036, "Python id": 5037, "finished": true, "Ev Idx": 7937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906288.352, "dur": 0.223,
    "args": {
      "Python parent id": 5037, "Python id": 5038, "finished": true, "Ev Idx": 7938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906288.444, "dur": 0.098,
    "args": {
      "Python parent id": 5038, "Python id": 5039, "finished": true, "Ev Idx": 7939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906288.873, "dur": 8.311,
    "args": {
      "Python parent id": 5036, "Python id": 5040, "finished": true, "Ev Idx": 7940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906289.017, "dur": 0.059,
    "args": {
      "Python parent id": 5040, "Python id": 5041, "finished": true, "Ev Idx": 7941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906289.166, "dur": 0.061,
    "args": {
      "Python parent id": 5040, "Python id": 5042, "finished": true, "Ev Idx": 7942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906289.358, "dur": 0.060,
    "args": {
      "Python parent id": 5040, "Python id": 5043, "finished": true, "Ev Idx": 7943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906289.800, "dur": 0.040,
    "args": {
      "Python parent id": 5040, "Python id": 5044, "finished": true, "Ev Idx": 7944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906290.149, "dur": 0.069,
    "args": {
      "Python parent id": 5040, "Python id": 5045, "finished": true, "Ev Idx": 7945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906290.578, "dur": 6.558,
    "args": {
      "Python parent id": 5040, "Python id": 5046, "finished": true, "Ev Idx": 7946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906290.707, "dur": 0.052,
    "args": {
      "Python parent id": 5046, "Python id": 5047, "finished": true, "Ev Idx": 7947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906290.939, "dur": 0.054,
    "args": {
      "Python parent id": 5046, "Python id": 5048, "finished": true, "Ev Idx": 7948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906291.170, "dur": 0.066,
    "args": {
      "Python parent id": 5046, "Python id": 5049, "finished": true, "Ev Idx": 7949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906291.565, "dur": 0.061,
    "args": {
      "Python parent id": 5046, "Python id": 5050, "finished": true, "Ev Idx": 7950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906291.925, "dur": 5.121,
    "args": {
      "Python parent id": 5046, "Python id": 5051, "finished": true, "Ev Idx": 7951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906292.272, "dur": 4.721,
    "args": {
      "Python parent id": 5051, "Python id": 5052, "finished": true, "Ev Idx": 7952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906292.391, "dur": 2.884,
    "args": {
      "Python parent id": 5052, "Python id": 5053, "finished": true, "Ev Idx": 7953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906292.573, "dur": 2.657,
    "args": {
      "Python parent id": 5053, "Python id": 5054, "finished": true, "Ev Idx": 7954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906292.696, "dur": 0.277,
    "args": {
      "Python parent id": 5054, "Python id": 5055, "finished": true, "Ev Idx": 7955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906292.817, "dur": 0.062,
    "args": {
      "Python parent id": 5055, "Python id": 5056, "finished": true, "Ev Idx": 7956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.122, "dur": 1.850,
    "args": {
      "Python parent id": 5054, "Python id": 5057, "finished": true, "Ev Idx": 7957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.277, "dur": 1.637,
    "args": {
      "Python parent id": 5057, "Python id": 5058, "finished": true, "Ev Idx": 7958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.425, "dur": 1.428,
    "args": {
      "Python parent id": 5058, "Python id": 5059, "finished": true, "Ev Idx": 7959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.551, "dur": 1.082,
    "args": {
      "Python parent id": 5059, "Python id": 5060, "finished": true, "Ev Idx": 7960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.739, "dur": 0.462,
    "args": {
      "Python parent id": 5060, "Python id": 5061, "finished": true, "Ev Idx": 7961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906293.821, "dur": 0.047,
    "args": {
      "Python parent id": 5061, "Python id": 5062, "finished": true, "Ev Idx": 7962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906294.015, "dur": 0.141,
    "args": {
      "Python parent id": 5061, "Python id": 5063, "finished": true, "Ev Idx": 7963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906295.140, "dur": 0.060,
    "args": {
      "Python parent id": 5054, "Python id": 5064, "finished": true, "Ev Idx": 7964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906295.451, "dur": 0.109,
    "args": {
      "Python parent id": 5052, "Python id": 5065, "finished": true, "Ev Idx": 7965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906295.829, "dur": 1.123,
    "args": {
      "Python parent id": 5052, "Python id": 5066, "finished": true, "Ev Idx": 7966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906295.981, "dur": 0.936,
    "args": {
      "Python parent id": 5066, "Python id": 5067, "finished": true, "Ev Idx": 7967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906296.150, "dur": 0.359,
    "args": {
      "Python parent id": 5067, "Python id": 5068, "finished": true, "Ev Idx": 7968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906296.282, "dur": 0.168,
    "args": {
      "Python parent id": 5068, "Python id": 5069, "finished": true, "Ev Idx": 7969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906296.361, "dur": 0.058,
    "args": {
      "Python parent id": 5069, "Python id": 5070, "finished": true, "Ev Idx": 7970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906296.633, "dur": 0.234,
    "args": {
      "Python parent id": 5067, "Python id": 5071, "finished": true, "Ev Idx": 7971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906297.318, "dur": 0.146,
    "args": {
      "Python parent id": 5036, "Python id": 5072, "finished": true, "Ev Idx": 7972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906298.024, "dur": 0.877,
    "args": {
      "Python parent id": 5036, "Python id": 5073, "finished": true, "Ev Idx": 7973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906298.411, "dur": 0.442,
    "args": {
      "Python parent id": 5073, "Python id": 5074, "finished": true, "Ev Idx": 7974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906299.294, "dur": 2.532,
    "args": {
      "Python parent id": 5035, "Python id": 5075, "finished": true, "Ev Idx": 7975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906302.820, "dur": 80.706,
    "args": {
      "Python parent id": 12, "Python id": 5076, "finished": true, "Ev Idx": 7976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906303.171, "dur": 0.450,
    "args": {
      "Python parent id": 5076, "Python id": 5077, "finished": true, "Ev Idx": 7977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906303.482, "dur": 0.094,
    "args": {
      "Python parent id": 5077, "Python id": 5078, "finished": true, "Ev Idx": 7978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906303.937, "dur": 0.413,
    "args": {
      "Python parent id": 5076, "Python id": 5079, "finished": true, "Ev Idx": 7979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906304.539, "dur": 0.253,
    "args": {
      "Python parent id": 5076, "Python id": 5080, "finished": true, "Ev Idx": 7980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906304.996, "dur": 1.580,
    "args": {
      "Python parent id": 5076, "Python id": 5081, "finished": true, "Ev Idx": 7981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906307.213, "dur": 76.245,
    "args": {
      "Python parent id": 5076, "Python id": 5082, "finished": true, "Ev Idx": 7982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906308.073, "dur": 75.185,
    "args": {
      "Python parent id": 5082, "Python id": 5083, "finished": true, "Ev Idx": 7983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906308.393, "dur": 74.788,
    "args": {
      "Python parent id": 5083, "Python id": 5084, "finished": true, "Ev Idx": 7984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906313.572, "dur": 68.285,
    "args": {
      "Python parent id": 5084, "Python id": 5085, "finished": true, "Ev Idx": 7985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906313.946, "dur": 67.812,
    "args": {
      "Python parent id": 5085, "Python id": 5086, "finished": true, "Ev Idx": 7986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906314.501, "dur": 67.099,
    "args": {
      "Python parent id": 5086, "Python id": 5087, "finished": true, "Ev Idx": 7987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906315.069, "dur": 0.443,
    "args": {
      "Python parent id": 5087, "Python id": 5088, "finished": true, "Ev Idx": 7988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906315.362, "dur": 0.080,
    "args": {
      "Python parent id": 5088, "Python id": 5089, "finished": true, "Ev Idx": 7989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906315.783, "dur": 0.314,
    "args": {
      "Python parent id": 5087, "Python id": 5090, "finished": true, "Ev Idx": 7990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906315.995, "dur": 0.048,
    "args": {
      "Python parent id": 5090, "Python id": 5091, "finished": true, "Ev Idx": 7991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906316.393, "dur": 65.081,
    "args": {
      "Python parent id": 5087, "Python id": 5092, "finished": true, "Ev Idx": 7992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906317.144, "dur": 63.181,
    "args": {
      "Python parent id": 5092, "Python id": 5093, "finished": true, "Ev Idx": 7993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906323.941, "dur": 16.337,
    "args": {
      "Python parent id": 5093, "Python id": 5094, "finished": true, "Ev Idx": 7994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906324.047, "dur": 0.124,
    "args": {
      "Python parent id": 5094, "Python id": 5095, "finished": true, "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906324.496, "dur": 15.717,
    "args": {
      "Python parent id": 5094, "Python id": 5096, "finished": true, "Ev Idx": 7996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906324.647, "dur": 7.139,
    "args": {
      "Python parent id": 5096, "Python id": 5097, "finished": true, "Ev Idx": 7997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906346.633, "dur": 6.126,
    "args": {
      "Python parent id": 5093, "Python id": 5098, "finished": true, "Ev Idx": 7998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906360.172, "dur": 7.033,
    "args": {
      "Python parent id": 5093, "Python id": 5099, "finished": true, "Ev Idx": 7999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906384.644, "dur": 17.203,
    "args": {
      "Python parent id": 12, "Python id": 5100, "finished": true, "Ev Idx": 8000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906385.058, "dur": 13.048,
    "args": {
      "Python parent id": 5100, "Python id": 5101, "finished": true, "Ev Idx": 8001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906385.255, "dur": 0.486,
    "args": {
      "Python parent id": 5101, "Python id": 5102, "finished": true, "Ev Idx": 8002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906385.408, "dur": 0.274,
    "args": {
      "Python parent id": 5102, "Python id": 5103, "finished": true, "Ev Idx": 8003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906385.581, "dur": 0.069,
    "args": {
      "Python parent id": 5103, "Python id": 5104, "finished": true, "Ev Idx": 8004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906386.096, "dur": 8.544,
    "args": {
      "Python parent id": 5101, "Python id": 5105, "finished": true, "Ev Idx": 8005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906386.205, "dur": 0.059,
    "args": {
      "Python parent id": 5105, "Python id": 5106, "finished": true, "Ev Idx": 8006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906386.358, "dur": 0.068,
    "args": {
      "Python parent id": 5105, "Python id": 5107, "finished": true, "Ev Idx": 8007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906386.561, "dur": 0.051,
    "args": {
      "Python parent id": 5105, "Python id": 5108, "finished": true, "Ev Idx": 8008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906387.084, "dur": 0.026,
    "args": {
      "Python parent id": 5105, "Python id": 5109, "finished": true, "Ev Idx": 8009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906387.361, "dur": 0.067,
    "args": {
      "Python parent id": 5105, "Python id": 5110, "finished": true, "Ev Idx": 8010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906387.795, "dur": 6.761,
    "args": {
      "Python parent id": 5105, "Python id": 5111, "finished": true, "Ev Idx": 8011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906387.902, "dur": 0.062,
    "args": {
      "Python parent id": 5111, "Python id": 5112, "finished": true, "Ev Idx": 8012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906388.123, "dur": 0.054,
    "args": {
      "Python parent id": 5111, "Python id": 5113, "finished": true, "Ev Idx": 8013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906388.324, "dur": 0.065,
    "args": {
      "Python parent id": 5111, "Python id": 5114, "finished": true, "Ev Idx": 8014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906388.683, "dur": 0.034,
    "args": {
      "Python parent id": 5111, "Python id": 5115, "finished": true, "Ev Idx": 8015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.042, "dur": 5.460,
    "args": {
      "Python parent id": 5111, "Python id": 5116, "finished": true, "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.378, "dur": 5.069,
    "args": {
      "Python parent id": 5116, "Python id": 5117, "finished": true, "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.514, "dur": 3.199,
    "args": {
      "Python parent id": 5117, "Python id": 5118, "finished": true, "Ev Idx": 8018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.686, "dur": 2.985,
    "args": {
      "Python parent id": 5118, "Python id": 5119, "finished": true, "Ev Idx": 8019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.807, "dur": 0.242,
    "args": {
      "Python parent id": 5119, "Python id": 5120, "finished": true, "Ev Idx": 8020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906389.913, "dur": 0.064,
    "args": {
      "Python parent id": 5120, "Python id": 5121, "finished": true, "Ev Idx": 8021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906390.199, "dur": 2.157,
    "args": {
      "Python parent id": 5119, "Python id": 5122, "finished": true, "Ev Idx": 8022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906390.385, "dur": 1.886,
    "args": {
      "Python parent id": 5122, "Python id": 5123, "finished": true, "Ev Idx": 8023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906390.554, "dur": 1.650,
    "args": {
      "Python parent id": 5123, "Python id": 5124, "finished": true, "Ev Idx": 8024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906390.744, "dur": 1.229,
    "args": {
      "Python parent id": 5124, "Python id": 5125, "finished": true, "Ev Idx": 8025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906390.952, "dur": 0.533,
    "args": {
      "Python parent id": 5125, "Python id": 5126, "finished": true, "Ev Idx": 8026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906391.031, "dur": 0.046,
    "args": {
      "Python parent id": 5126, "Python id": 5127, "finished": true, "Ev Idx": 8027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906391.257, "dur": 0.185,
    "args": {
      "Python parent id": 5126, "Python id": 5128, "finished": true, "Ev Idx": 8028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906392.533, "dur": 0.108,
    "args": {
      "Python parent id": 5119, "Python id": 5129, "finished": true, "Ev Idx": 8029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906392.884, "dur": 0.116,
    "args": {
      "Python parent id": 5117, "Python id": 5130, "finished": true, "Ev Idx": 8030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906393.293, "dur": 1.116,
    "args": {
      "Python parent id": 5117, "Python id": 5131, "finished": true, "Ev Idx": 8031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906393.439, "dur": 0.931,
    "args": {
      "Python parent id": 5131, "Python id": 5132, "finished": true, "Ev Idx": 8032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906393.585, "dur": 0.376,
    "args": {
      "Python parent id": 5132, "Python id": 5133, "finished": true, "Ev Idx": 8033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906393.712, "dur": 0.192,
    "args": {
      "Python parent id": 5133, "Python id": 5134, "finished": true, "Ev Idx": 8034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906393.827, "dur": 0.050,
    "args": {
      "Python parent id": 5134, "Python id": 5135, "finished": true, "Ev Idx": 8035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906394.078, "dur": 0.260,
    "args": {
      "Python parent id": 5132, "Python id": 5136, "finished": true, "Ev Idx": 8036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906394.777, "dur": 0.234,
    "args": {
      "Python parent id": 5101, "Python id": 5137, "finished": true, "Ev Idx": 8037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906395.652, "dur": 2.302,
    "args": {
      "Python parent id": 5101, "Python id": 5138, "finished": true, "Ev Idx": 8038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906397.378, "dur": 0.523,
    "args": {
      "Python parent id": 5138, "Python id": 5139, "finished": true, "Ev Idx": 8039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906398.441, "dur": 3.326,
    "args": {
      "Python parent id": 5100, "Python id": 5140, "finished": true, "Ev Idx": 8040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906402.491, "dur": 6.684,
    "args": {
      "Python parent id": 12, "Python id": 5141, "finished": true, "Ev Idx": 8041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906409.655, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 5142, "finished": true, "Ev Idx": 8042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906410.366, "dur": 0.218,
    "args": {
      "Python parent id": 12, "Python id": 5143, "finished": true, "Ev Idx": 8043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906411.069, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 5144, "finished": true, "Ev Idx": 8044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906411.504, "dur": 11.200,
    "args": {
      "Python parent id": 12, "Python id": 5145, "finished": true, "Ev Idx": 8045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906423.935, "dur": 14.185,
    "args": {
      "Python parent id": 12, "Python id": 5146, "finished": true, "Ev Idx": 8046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906424.254, "dur": 10.880,
    "args": {
      "Python parent id": 5146, "Python id": 5147, "finished": true, "Ev Idx": 8047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906424.398, "dur": 0.522,
    "args": {
      "Python parent id": 5147, "Python id": 5148, "finished": true, "Ev Idx": 8048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906424.597, "dur": 0.229,
    "args": {
      "Python parent id": 5148, "Python id": 5149, "finished": true, "Ev Idx": 8049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906424.709, "dur": 0.086,
    "args": {
      "Python parent id": 5149, "Python id": 5150, "finished": true, "Ev Idx": 8050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906425.295, "dur": 8.070,
    "args": {
      "Python parent id": 5147, "Python id": 5151, "finished": true, "Ev Idx": 8051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906425.411, "dur": 0.079,
    "args": {
      "Python parent id": 5151, "Python id": 5152, "finished": true, "Ev Idx": 8052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906425.603, "dur": 0.054,
    "args": {
      "Python parent id": 5151, "Python id": 5153, "finished": true, "Ev Idx": 8053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906425.830, "dur": 0.060,
    "args": {
      "Python parent id": 5151, "Python id": 5154, "finished": true, "Ev Idx": 8054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906426.326, "dur": 0.036,
    "args": {
      "Python parent id": 5151, "Python id": 5155, "finished": true, "Ev Idx": 8055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906426.621, "dur": 0.073,
    "args": {
      "Python parent id": 5151, "Python id": 5156, "finished": true, "Ev Idx": 8056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906427.034, "dur": 6.279,
    "args": {
      "Python parent id": 5151, "Python id": 5157, "finished": true, "Ev Idx": 8057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906427.131, "dur": 0.056,
    "args": {
      "Python parent id": 5157, "Python id": 5158, "finished": true, "Ev Idx": 8058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906427.319, "dur": 0.063,
    "args": {
      "Python parent id": 5157, "Python id": 5159, "finished": true, "Ev Idx": 8059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906427.572, "dur": 0.063,
    "args": {
      "Python parent id": 5157, "Python id": 5160, "finished": true, "Ev Idx": 8060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906427.993, "dur": 0.034,
    "args": {
      "Python parent id": 5157, "Python id": 5161, "finished": true, "Ev Idx": 8061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906428.316, "dur": 4.911,
    "args": {
      "Python parent id": 5157, "Python id": 5162, "finished": true, "Ev Idx": 8062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906428.495, "dur": 4.679,
    "args": {
      "Python parent id": 5162, "Python id": 5163, "finished": true, "Ev Idx": 8063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906428.625, "dur": 2.973,
    "args": {
      "Python parent id": 5163, "Python id": 5164, "finished": true, "Ev Idx": 8064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906428.823, "dur": 2.732,
    "args": {
      "Python parent id": 5164, "Python id": 5165, "finished": true, "Ev Idx": 8065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906428.959, "dur": 0.252,
    "args": {
      "Python parent id": 5165, "Python id": 5166, "finished": true, "Ev Idx": 8066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906429.075, "dur": 0.064,
    "args": {
      "Python parent id": 5166, "Python id": 5167, "finished": true, "Ev Idx": 8067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906429.347, "dur": 1.944,
    "args": {
      "Python parent id": 5165, "Python id": 5168, "finished": true, "Ev Idx": 8068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906429.532, "dur": 1.703,
    "args": {
      "Python parent id": 5168, "Python id": 5169, "finished": true, "Ev Idx": 8069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906429.705, "dur": 1.461,
    "args": {
      "Python parent id": 5169, "Python id": 5170, "finished": true, "Ev Idx": 8070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906429.855, "dur": 1.080,
    "args": {
      "Python parent id": 5170, "Python id": 5171, "finished": true, "Ev Idx": 8071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906430.038, "dur": 0.467,
    "args": {
      "Python parent id": 5171, "Python id": 5172, "finished": true, "Ev Idx": 8072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906430.114, "dur": 0.054,
    "args": {
      "Python parent id": 5172, "Python id": 5173, "finished": true, "Ev Idx": 8073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906430.332, "dur": 0.132,
    "args": {
      "Python parent id": 5172, "Python id": 5174, "finished": true, "Ev Idx": 8074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906431.459, "dur": 0.064,
    "args": {
      "Python parent id": 5165, "Python id": 5175, "finished": true, "Ev Idx": 8075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906431.753, "dur": 0.068,
    "args": {
      "Python parent id": 5163, "Python id": 5176, "finished": true, "Ev Idx": 8076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.038, "dur": 1.095,
    "args": {
      "Python parent id": 5163, "Python id": 5177, "finished": true, "Ev Idx": 8077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.182, "dur": 0.911,
    "args": {
      "Python parent id": 5177, "Python id": 5178, "finished": true, "Ev Idx": 8078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.305, "dur": 0.378,
    "args": {
      "Python parent id": 5178, "Python id": 5179, "finished": true, "Ev Idx": 8079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.442, "dur": 0.182,
    "args": {
      "Python parent id": 5179, "Python id": 5180, "finished": true, "Ev Idx": 8080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.544, "dur": 0.048,
    "args": {
      "Python parent id": 5180, "Python id": 5181, "finished": true, "Ev Idx": 8081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906432.810, "dur": 0.253,
    "args": {
      "Python parent id": 5178, "Python id": 5182, "finished": true, "Ev Idx": 8082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906433.479, "dur": 0.189,
    "args": {
      "Python parent id": 5147, "Python id": 5183, "finished": true, "Ev Idx": 8083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906434.166, "dur": 0.813,
    "args": {
      "Python parent id": 5147, "Python id": 5184, "finished": true, "Ev Idx": 8084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906434.530, "dur": 0.401,
    "args": {
      "Python parent id": 5184, "Python id": 5185, "finished": true, "Ev Idx": 8085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906435.393, "dur": 2.684,
    "args": {
      "Python parent id": 5146, "Python id": 5186, "finished": true, "Ev Idx": 8086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906439.021, "dur": 92.845,
    "args": {
      "Python parent id": 12, "Python id": 5187, "finished": true, "Ev Idx": 8087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906439.375, "dur": 0.360,
    "args": {
      "Python parent id": 5187, "Python id": 5188, "finished": true, "Ev Idx": 8088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906439.613, "dur": 0.077,
    "args": {
      "Python parent id": 5188, "Python id": 5189, "finished": true, "Ev Idx": 8089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906440.044, "dur": 0.422,
    "args": {
      "Python parent id": 5187, "Python id": 5190, "finished": true, "Ev Idx": 8090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906440.690, "dur": 0.250,
    "args": {
      "Python parent id": 5187, "Python id": 5191, "finished": true, "Ev Idx": 8091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906441.154, "dur": 1.590,
    "args": {
      "Python parent id": 5187, "Python id": 5192, "finished": true, "Ev Idx": 8092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906443.351, "dur": 88.387,
    "args": {
      "Python parent id": 5187, "Python id": 5193, "finished": true, "Ev Idx": 8093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906444.220, "dur": 87.351,
    "args": {
      "Python parent id": 5193, "Python id": 5194, "finished": true, "Ev Idx": 8094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906444.470, "dur": 87.042,
    "args": {
      "Python parent id": 5194, "Python id": 5195, "finished": true, "Ev Idx": 8095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906450.265, "dur": 79.954,
    "args": {
      "Python parent id": 5195, "Python id": 5196, "finished": true, "Ev Idx": 8096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906450.572, "dur": 79.520,
    "args": {
      "Python parent id": 5196, "Python id": 5197, "finished": true, "Ev Idx": 8097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906451.146, "dur": 78.767,
    "args": {
      "Python parent id": 5197, "Python id": 5198, "finished": true, "Ev Idx": 8098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906451.695, "dur": 0.387,
    "args": {
      "Python parent id": 5198, "Python id": 5199, "finished": true, "Ev Idx": 8099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906451.965, "dur": 0.072,
    "args": {
      "Python parent id": 5199, "Python id": 5200, "finished": true, "Ev Idx": 8100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906452.393, "dur": 0.352,
    "args": {
      "Python parent id": 5198, "Python id": 5201, "finished": true, "Ev Idx": 8101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906452.644, "dur": 0.053,
    "args": {
      "Python parent id": 5201, "Python id": 5202, "finished": true, "Ev Idx": 8102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906453.131, "dur": 76.675,
    "args": {
      "Python parent id": 5198, "Python id": 5203, "finished": true, "Ev Idx": 8103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906453.809, "dur": 74.705,
    "args": {
      "Python parent id": 5203, "Python id": 5204, "finished": true, "Ev Idx": 8104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906460.710, "dur": 21.151,
    "args": {
      "Python parent id": 5204, "Python id": 5205, "finished": true, "Ev Idx": 8105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906460.846, "dur": 0.071,
    "args": {
      "Python parent id": 5205, "Python id": 5206, "finished": true, "Ev Idx": 8106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906461.210, "dur": 20.547,
    "args": {
      "Python parent id": 5205, "Python id": 5207, "finished": true, "Ev Idx": 8107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906461.358, "dur": 7.278,
    "args": {
      "Python parent id": 5207, "Python id": 5208, "finished": true, "Ev Idx": 8108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906488.747, "dur": 6.652,
    "args": {
      "Python parent id": 5204, "Python id": 5209, "finished": true, "Ev Idx": 8109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906503.103, "dur": 11.820,
    "args": {
      "Python parent id": 5204, "Python id": 5210, "finished": true, "Ev Idx": 8110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906532.937, "dur": 18.051,
    "args": {
      "Python parent id": 12, "Python id": 5211, "finished": true, "Ev Idx": 8111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906533.345, "dur": 12.377,
    "args": {
      "Python parent id": 5211, "Python id": 5212, "finished": true, "Ev Idx": 8112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906533.505, "dur": 0.490,
    "args": {
      "Python parent id": 5212, "Python id": 5213, "finished": true, "Ev Idx": 8113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906533.679, "dur": 0.242,
    "args": {
      "Python parent id": 5213, "Python id": 5214, "finished": true, "Ev Idx": 8114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906533.774, "dur": 0.086,
    "args": {
      "Python parent id": 5214, "Python id": 5215, "finished": true, "Ev Idx": 8115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906534.340, "dur": 9.381,
    "args": {
      "Python parent id": 5212, "Python id": 5216, "finished": true, "Ev Idx": 8116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906534.459, "dur": 0.129,
    "args": {
      "Python parent id": 5216, "Python id": 5217, "finished": true, "Ev Idx": 8117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906534.675, "dur": 0.062,
    "args": {
      "Python parent id": 5216, "Python id": 5218, "finished": true, "Ev Idx": 8118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906534.863, "dur": 0.062,
    "args": {
      "Python parent id": 5216, "Python id": 5219, "finished": true, "Ev Idx": 8119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906535.354, "dur": 0.041,
    "args": {
      "Python parent id": 5216, "Python id": 5220, "finished": true, "Ev Idx": 8120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906535.663, "dur": 0.073,
    "args": {
      "Python parent id": 5216, "Python id": 5221, "finished": true, "Ev Idx": 8121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906536.097, "dur": 7.570,
    "args": {
      "Python parent id": 5216, "Python id": 5222, "finished": true, "Ev Idx": 8122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906536.201, "dur": 0.063,
    "args": {
      "Python parent id": 5222, "Python id": 5223, "finished": true, "Ev Idx": 8123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906536.405, "dur": 0.058,
    "args": {
      "Python parent id": 5222, "Python id": 5224, "finished": true, "Ev Idx": 8124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906536.632, "dur": 0.071,
    "args": {
      "Python parent id": 5222, "Python id": 5225, "finished": true, "Ev Idx": 8125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906536.991, "dur": 0.038,
    "args": {
      "Python parent id": 5222, "Python id": 5226, "finished": true, "Ev Idx": 8126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906537.303, "dur": 6.259,
    "args": {
      "Python parent id": 5222, "Python id": 5227, "finished": true, "Ev Idx": 8127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906537.545, "dur": 5.964,
    "args": {
      "Python parent id": 5227, "Python id": 5228, "finished": true, "Ev Idx": 8128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906537.690, "dur": 3.229,
    "args": {
      "Python parent id": 5228, "Python id": 5229, "finished": true, "Ev Idx": 8129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906537.860, "dur": 3.017,
    "args": {
      "Python parent id": 5229, "Python id": 5230, "finished": true, "Ev Idx": 8130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.029, "dur": 0.259,
    "args": {
      "Python parent id": 5230, "Python id": 5231, "finished": true, "Ev Idx": 8131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.150, "dur": 0.068,
    "args": {
      "Python parent id": 5231, "Python id": 5232, "finished": true, "Ev Idx": 8132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.450, "dur": 2.077,
    "args": {
      "Python parent id": 5230, "Python id": 5233, "finished": true, "Ev Idx": 8133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.633, "dur": 1.820,
    "args": {
      "Python parent id": 5233, "Python id": 5234, "finished": true, "Ev Idx": 8134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.805, "dur": 1.580,
    "args": {
      "Python parent id": 5234, "Python id": 5235, "finished": true, "Ev Idx": 8135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906538.976, "dur": 1.176,
    "args": {
      "Python parent id": 5235, "Python id": 5236, "finished": true, "Ev Idx": 8136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906539.143, "dur": 0.470,
    "args": {
      "Python parent id": 5236, "Python id": 5237, "finished": true, "Ev Idx": 8137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906539.223, "dur": 0.060,
    "args": {
      "Python parent id": 5237, "Python id": 5238, "finished": true, "Ev Idx": 8138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906539.452, "dur": 0.119,
    "args": {
      "Python parent id": 5237, "Python id": 5239, "finished": true, "Ev Idx": 8139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906540.707, "dur": 0.140,
    "args": {
      "Python parent id": 5230, "Python id": 5240, "finished": true, "Ev Idx": 8140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906541.973, "dur": 0.110,
    "args": {
      "Python parent id": 5228, "Python id": 5241, "finished": true, "Ev Idx": 8141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906542.360, "dur": 1.111,
    "args": {
      "Python parent id": 5228, "Python id": 5242, "finished": true, "Ev Idx": 8142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906542.508, "dur": 0.919,
    "args": {
      "Python parent id": 5242, "Python id": 5243, "finished": true, "Ev Idx": 8143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906542.646, "dur": 0.379,
    "args": {
      "Python parent id": 5243, "Python id": 5244, "finished": true, "Ev Idx": 8144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906542.790, "dur": 0.174,
    "args": {
      "Python parent id": 5244, "Python id": 5245, "finished": true, "Ev Idx": 8145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906542.875, "dur": 0.057,
    "args": {
      "Python parent id": 5245, "Python id": 5246, "finished": true, "Ev Idx": 8146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906543.143, "dur": 0.252,
    "args": {
      "Python parent id": 5243, "Python id": 5247, "finished": true, "Ev Idx": 8147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906543.855, "dur": 0.217,
    "args": {
      "Python parent id": 5212, "Python id": 5248, "finished": true, "Ev Idx": 8148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906544.593, "dur": 0.988,
    "args": {
      "Python parent id": 5212, "Python id": 5249, "finished": true, "Ev Idx": 8149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906545.038, "dur": 0.495,
    "args": {
      "Python parent id": 5249, "Python id": 5250, "finished": true, "Ev Idx": 8150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906545.981, "dur": 4.913,
    "args": {
      "Python parent id": 5211, "Python id": 5251, "finished": true, "Ev Idx": 8151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906551.607, "dur": 7.966,
    "args": {
      "Python parent id": 12, "Python id": 5252, "finished": true, "Ev Idx": 8152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906560.057, "dur": 0.091,
    "args": {
      "Python parent id": 12, "Python id": 5253, "finished": true, "Ev Idx": 8153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906560.663, "dur": 0.194,
    "args": {
      "Python parent id": 12, "Python id": 5254, "finished": true, "Ev Idx": 8154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906561.341, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 5255, "finished": true, "Ev Idx": 8155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906561.787, "dur": 9.980,
    "args": {
      "Python parent id": 12, "Python id": 5256, "finished": true, "Ev Idx": 8156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906572.991, "dur": 14.277,
    "args": {
      "Python parent id": 12, "Python id": 5257, "finished": true, "Ev Idx": 8157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906573.364, "dur": 10.490,
    "args": {
      "Python parent id": 5257, "Python id": 5258, "finished": true, "Ev Idx": 8158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906573.530, "dur": 0.441,
    "args": {
      "Python parent id": 5258, "Python id": 5259, "finished": true, "Ev Idx": 8159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906573.692, "dur": 0.204,
    "args": {
      "Python parent id": 5259, "Python id": 5260, "finished": true, "Ev Idx": 8160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906573.774, "dur": 0.088,
    "args": {
      "Python parent id": 5260, "Python id": 5261, "finished": true, "Ev Idx": 8161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906574.197, "dur": 7.928,
    "args": {
      "Python parent id": 5258, "Python id": 5262, "finished": true, "Ev Idx": 8162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906574.279, "dur": 0.067,
    "args": {
      "Python parent id": 5262, "Python id": 5263, "finished": true, "Ev Idx": 8163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906574.456, "dur": 0.060,
    "args": {
      "Python parent id": 5262, "Python id": 5264, "finished": true, "Ev Idx": 8164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906574.680, "dur": 0.062,
    "args": {
      "Python parent id": 5262, "Python id": 5265, "finished": true, "Ev Idx": 8165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906575.209, "dur": 0.077,
    "args": {
      "Python parent id": 5262, "Python id": 5266, "finished": true, "Ev Idx": 8166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906575.613, "dur": 0.073,
    "args": {
      "Python parent id": 5262, "Python id": 5267, "finished": true, "Ev Idx": 8167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906576.085, "dur": 5.978,
    "args": {
      "Python parent id": 5262, "Python id": 5268, "finished": true, "Ev Idx": 8168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906576.200, "dur": 0.061,
    "args": {
      "Python parent id": 5268, "Python id": 5269, "finished": true, "Ev Idx": 8169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906576.396, "dur": 0.056,
    "args": {
      "Python parent id": 5268, "Python id": 5270, "finished": true, "Ev Idx": 8170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906576.591, "dur": 0.068,
    "args": {
      "Python parent id": 5268, "Python id": 5271, "finished": true, "Ev Idx": 8171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906576.937, "dur": 0.037,
    "args": {
      "Python parent id": 5268, "Python id": 5272, "finished": true, "Ev Idx": 8172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906577.291, "dur": 4.700,
    "args": {
      "Python parent id": 5268, "Python id": 5273, "finished": true, "Ev Idx": 8173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906577.495, "dur": 4.443,
    "args": {
      "Python parent id": 5273, "Python id": 5274, "finished": true, "Ev Idx": 8174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906577.611, "dur": 2.784,
    "args": {
      "Python parent id": 5274, "Python id": 5275, "finished": true, "Ev Idx": 8175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906577.786, "dur": 2.566,
    "args": {
      "Python parent id": 5275, "Python id": 5276, "finished": true, "Ev Idx": 8176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906577.937, "dur": 0.219,
    "args": {
      "Python parent id": 5276, "Python id": 5277, "finished": true, "Ev Idx": 8177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.037, "dur": 0.069,
    "args": {
      "Python parent id": 5277, "Python id": 5278, "finished": true, "Ev Idx": 8178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.316, "dur": 1.743,
    "args": {
      "Python parent id": 5276, "Python id": 5279, "finished": true, "Ev Idx": 8179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.466, "dur": 1.534,
    "args": {
      "Python parent id": 5279, "Python id": 5280, "finished": true, "Ev Idx": 8180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.659, "dur": 1.281,
    "args": {
      "Python parent id": 5280, "Python id": 5281, "finished": true, "Ev Idx": 8181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.777, "dur": 0.960,
    "args": {
      "Python parent id": 5281, "Python id": 5282, "finished": true, "Ev Idx": 8182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906578.929, "dur": 0.439,
    "args": {
      "Python parent id": 5282, "Python id": 5283, "finished": true, "Ev Idx": 8183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906579.006, "dur": 0.052,
    "args": {
      "Python parent id": 5283, "Python id": 5284, "finished": true, "Ev Idx": 8184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906579.207, "dur": 0.120,
    "args": {
      "Python parent id": 5283, "Python id": 5285, "finished": true, "Ev Idx": 8185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906580.224, "dur": 0.094,
    "args": {
      "Python parent id": 5276, "Python id": 5286, "finished": true, "Ev Idx": 8186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906580.544, "dur": 0.072,
    "args": {
      "Python parent id": 5274, "Python id": 5287, "finished": true, "Ev Idx": 8187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906580.864, "dur": 1.032,
    "args": {
      "Python parent id": 5274, "Python id": 5288, "finished": true, "Ev Idx": 8188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906580.994, "dur": 0.864,
    "args": {
      "Python parent id": 5288, "Python id": 5289, "finished": true, "Ev Idx": 8189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906581.125, "dur": 0.348,
    "args": {
      "Python parent id": 5289, "Python id": 5290, "finished": true, "Ev Idx": 8190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906581.259, "dur": 0.160,
    "args": {
      "Python parent id": 5290, "Python id": 5291, "finished": true, "Ev Idx": 8191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906581.335, "dur": 0.056,
    "args": {
      "Python parent id": 5291, "Python id": 5292, "finished": true, "Ev Idx": 8192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906581.599, "dur": 0.229,
    "args": {
      "Python parent id": 5289, "Python id": 5293, "finished": true, "Ev Idx": 8193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906582.240, "dur": 0.148,
    "args": {
      "Python parent id": 5258, "Python id": 5294, "finished": true, "Ev Idx": 8194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906582.931, "dur": 0.809,
    "args": {
      "Python parent id": 5258, "Python id": 5295, "finished": true, "Ev Idx": 8195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906583.317, "dur": 0.374,
    "args": {
      "Python parent id": 5295, "Python id": 5296, "finished": true, "Ev Idx": 8196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906584.074, "dur": 3.151,
    "args": {
      "Python parent id": 5257, "Python id": 5297, "finished": true, "Ev Idx": 8197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906588.151, "dur": 85.777,
    "args": {
      "Python parent id": 12, "Python id": 5298, "finished": true, "Ev Idx": 8198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906588.552, "dur": 0.381,
    "args": {
      "Python parent id": 5298, "Python id": 5299, "finished": true, "Ev Idx": 8199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906588.798, "dur": 0.088,
    "args": {
      "Python parent id": 5299, "Python id": 5300, "finished": true, "Ev Idx": 8200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906589.210, "dur": 0.317,
    "args": {
      "Python parent id": 5298, "Python id": 5301, "finished": true, "Ev Idx": 8201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906589.754, "dur": 0.247,
    "args": {
      "Python parent id": 5298, "Python id": 5302, "finished": true, "Ev Idx": 8202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906590.199, "dur": 1.515,
    "args": {
      "Python parent id": 5298, "Python id": 5303, "finished": true, "Ev Idx": 8203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906592.285, "dur": 81.580,
    "args": {
      "Python parent id": 5298, "Python id": 5304, "finished": true, "Ev Idx": 8204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906593.155, "dur": 80.505,
    "args": {
      "Python parent id": 5304, "Python id": 5305, "finished": true, "Ev Idx": 8205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906593.404, "dur": 80.202,
    "args": {
      "Python parent id": 5305, "Python id": 5306, "finished": true, "Ev Idx": 8206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906598.616, "dur": 73.682,
    "args": {
      "Python parent id": 5306, "Python id": 5307, "finished": true, "Ev Idx": 8207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906598.964, "dur": 73.242,
    "args": {
      "Python parent id": 5307, "Python id": 5308, "finished": true, "Ev Idx": 8208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906599.469, "dur": 72.580,
    "args": {
      "Python parent id": 5308, "Python id": 5309, "finished": true, "Ev Idx": 8209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906599.963, "dur": 0.449,
    "args": {
      "Python parent id": 5309, "Python id": 5310, "finished": true, "Ev Idx": 8210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906600.257, "dur": 0.092,
    "args": {
      "Python parent id": 5310, "Python id": 5311, "finished": true, "Ev Idx": 8211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906600.733, "dur": 0.322,
    "args": {
      "Python parent id": 5309, "Python id": 5312, "finished": true, "Ev Idx": 8212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906600.964, "dur": 0.042,
    "args": {
      "Python parent id": 5312, "Python id": 5313, "finished": true, "Ev Idx": 8213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906601.413, "dur": 70.550,
    "args": {
      "Python parent id": 5309, "Python id": 5314, "finished": true, "Ev Idx": 8214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906602.051, "dur": 68.740,
    "args": {
      "Python parent id": 5314, "Python id": 5315, "finished": true, "Ev Idx": 8215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906608.927, "dur": 16.289,
    "args": {
      "Python parent id": 5315, "Python id": 5316, "finished": true, "Ev Idx": 8216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906609.061, "dur": 0.115,
    "args": {
      "Python parent id": 5316, "Python id": 5317, "finished": true, "Ev Idx": 8217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906609.463, "dur": 15.650,
    "args": {
      "Python parent id": 5316, "Python id": 5318, "finished": true, "Ev Idx": 8218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906609.627, "dur": 7.001,
    "args": {
      "Python parent id": 5318, "Python id": 5319, "finished": true, "Ev Idx": 8219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906633.831, "dur": 6.901,
    "args": {
      "Python parent id": 5315, "Python id": 5320, "finished": true, "Ev Idx": 8220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906648.295, "dur": 8.618,
    "args": {
      "Python parent id": 5315, "Python id": 5321, "finished": true, "Ev Idx": 8221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906675.135, "dur": 14.823,
    "args": {
      "Python parent id": 12, "Python id": 5322, "finished": true, "Ev Idx": 8222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906675.505, "dur": 11.364,
    "args": {
      "Python parent id": 5322, "Python id": 5323, "finished": true, "Ev Idx": 8223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906675.680, "dur": 0.407,
    "args": {
      "Python parent id": 5323, "Python id": 5324, "finished": true, "Ev Idx": 8224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906675.827, "dur": 0.195,
    "args": {
      "Python parent id": 5324, "Python id": 5325, "finished": true, "Ev Idx": 8225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906675.921, "dur": 0.061,
    "args": {
      "Python parent id": 5325, "Python id": 5326, "finished": true, "Ev Idx": 8226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906676.422, "dur": 8.379,
    "args": {
      "Python parent id": 5323, "Python id": 5327, "finished": true, "Ev Idx": 8227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906676.534, "dur": 0.072,
    "args": {
      "Python parent id": 5327, "Python id": 5328, "finished": true, "Ev Idx": 8228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906676.697, "dur": 0.056,
    "args": {
      "Python parent id": 5327, "Python id": 5329, "finished": true, "Ev Idx": 8229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906676.924, "dur": 0.059,
    "args": {
      "Python parent id": 5327, "Python id": 5330, "finished": true, "Ev Idx": 8230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906677.429, "dur": 0.036,
    "args": {
      "Python parent id": 5327, "Python id": 5331, "finished": true, "Ev Idx": 8231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906677.727, "dur": 0.077,
    "args": {
      "Python parent id": 5327, "Python id": 5332, "finished": true, "Ev Idx": 8232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906678.147, "dur": 6.587,
    "args": {
      "Python parent id": 5327, "Python id": 5333, "finished": true, "Ev Idx": 8233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906678.240, "dur": 0.058,
    "args": {
      "Python parent id": 5333, "Python id": 5334, "finished": true, "Ev Idx": 8234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906678.438, "dur": 0.059,
    "args": {
      "Python parent id": 5333, "Python id": 5335, "finished": true, "Ev Idx": 8235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906678.683, "dur": 0.093,
    "args": {
      "Python parent id": 5333, "Python id": 5336, "finished": true, "Ev Idx": 8236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906679.124, "dur": 0.039,
    "args": {
      "Python parent id": 5333, "Python id": 5337, "finished": true, "Ev Idx": 8237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906679.462, "dur": 5.216,
    "args": {
      "Python parent id": 5333, "Python id": 5338, "finished": true, "Ev Idx": 8238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906679.691, "dur": 4.930,
    "args": {
      "Python parent id": 5338, "Python id": 5339, "finished": true, "Ev Idx": 8239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906679.817, "dur": 3.172,
    "args": {
      "Python parent id": 5339, "Python id": 5340, "finished": true, "Ev Idx": 8240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.024, "dur": 2.923,
    "args": {
      "Python parent id": 5340, "Python id": 5341, "finished": true, "Ev Idx": 8241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.163, "dur": 0.230,
    "args": {
      "Python parent id": 5341, "Python id": 5342, "finished": true, "Ev Idx": 8242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.280, "dur": 0.061,
    "args": {
      "Python parent id": 5342, "Python id": 5343, "finished": true, "Ev Idx": 8243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.556, "dur": 2.045,
    "args": {
      "Python parent id": 5341, "Python id": 5344, "finished": true, "Ev Idx": 8244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.795, "dur": 1.731,
    "args": {
      "Python parent id": 5344, "Python id": 5345, "finished": true, "Ev Idx": 8245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906680.974, "dur": 1.464,
    "args": {
      "Python parent id": 5345, "Python id": 5346, "finished": true, "Ev Idx": 8246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906681.149, "dur": 1.068,
    "args": {
      "Python parent id": 5346, "Python id": 5347, "finished": true, "Ev Idx": 8247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906681.306, "dur": 0.474,
    "args": {
      "Python parent id": 5347, "Python id": 5348, "finished": true, "Ev Idx": 8248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906681.401, "dur": 0.052,
    "args": {
      "Python parent id": 5348, "Python id": 5349, "finished": true, "Ev Idx": 8249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906681.598, "dur": 0.143,
    "args": {
      "Python parent id": 5348, "Python id": 5350, "finished": true, "Ev Idx": 8250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906682.758, "dur": 0.142,
    "args": {
      "Python parent id": 5341, "Python id": 5351, "finished": true, "Ev Idx": 8251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906683.145, "dur": 0.117,
    "args": {
      "Python parent id": 5339, "Python id": 5352, "finished": true, "Ev Idx": 8252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906683.540, "dur": 1.041,
    "args": {
      "Python parent id": 5339, "Python id": 5353, "finished": true, "Ev Idx": 8253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906683.679, "dur": 0.861,
    "args": {
      "Python parent id": 5353, "Python id": 5354, "finished": true, "Ev Idx": 8254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906683.822, "dur": 0.354,
    "args": {
      "Python parent id": 5354, "Python id": 5355, "finished": true, "Ev Idx": 8255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906683.953, "dur": 0.169,
    "args": {
      "Python parent id": 5355, "Python id": 5356, "finished": true, "Ev Idx": 8256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906684.039, "dur": 0.052,
    "args": {
      "Python parent id": 5356, "Python id": 5357, "finished": true, "Ev Idx": 8257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906684.300, "dur": 0.206,
    "args": {
      "Python parent id": 5354, "Python id": 5358, "finished": true, "Ev Idx": 8258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906684.935, "dur": 0.237,
    "args": {
      "Python parent id": 5323, "Python id": 5359, "finished": true, "Ev Idx": 8259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906685.697, "dur": 1.045,
    "args": {
      "Python parent id": 5323, "Python id": 5360, "finished": true, "Ev Idx": 8260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906686.177, "dur": 0.518,
    "args": {
      "Python parent id": 5360, "Python id": 5361, "finished": true, "Ev Idx": 8261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906687.149, "dur": 2.717,
    "args": {
      "Python parent id": 5322, "Python id": 5362, "finished": true, "Ev Idx": 8262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906690.475, "dur": 7.074,
    "args": {
      "Python parent id": 12, "Python id": 5363, "finished": true, "Ev Idx": 8263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906698.003, "dur": 0.126,
    "args": {
      "Python parent id": 12, "Python id": 5364, "finished": true, "Ev Idx": 8264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906698.710, "dur": 0.174,
    "args": {
      "Python parent id": 12, "Python id": 5365, "finished": true, "Ev Idx": 8265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906699.339, "dur": 0.083,
    "args": {
      "Python parent id": 12, "Python id": 5366, "finished": true, "Ev Idx": 8266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906699.793, "dur": 9.046,
    "args": {
      "Python parent id": 12, "Python id": 5367, "finished": true, "Ev Idx": 8267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906710.085, "dur": 13.957,
    "args": {
      "Python parent id": 12, "Python id": 5368, "finished": true, "Ev Idx": 8268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906710.429, "dur": 10.529,
    "args": {
      "Python parent id": 5368, "Python id": 5369, "finished": true, "Ev Idx": 8269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906710.564, "dur": 0.450,
    "args": {
      "Python parent id": 5369, "Python id": 5370, "finished": true, "Ev Idx": 8270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906710.708, "dur": 0.218,
    "args": {
      "Python parent id": 5370, "Python id": 5371, "finished": true, "Ev Idx": 8271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906710.801, "dur": 0.091,
    "args": {
      "Python parent id": 5371, "Python id": 5372, "finished": true, "Ev Idx": 8272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906711.235, "dur": 7.898,
    "args": {
      "Python parent id": 5369, "Python id": 5373, "finished": true, "Ev Idx": 8273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906711.316, "dur": 0.082,
    "args": {
      "Python parent id": 5373, "Python id": 5374, "finished": true, "Ev Idx": 8274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906711.523, "dur": 0.068,
    "args": {
      "Python parent id": 5373, "Python id": 5375, "finished": true, "Ev Idx": 8275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906711.722, "dur": 0.057,
    "args": {
      "Python parent id": 5373, "Python id": 5376, "finished": true, "Ev Idx": 8276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906712.166, "dur": 0.036,
    "args": {
      "Python parent id": 5373, "Python id": 5377, "finished": true, "Ev Idx": 8277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906712.463, "dur": 0.069,
    "args": {
      "Python parent id": 5373, "Python id": 5378, "finished": true, "Ev Idx": 8278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906712.871, "dur": 6.211,
    "args": {
      "Python parent id": 5373, "Python id": 5379, "finished": true, "Ev Idx": 8279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906712.952, "dur": 0.057,
    "args": {
      "Python parent id": 5379, "Python id": 5380, "finished": true, "Ev Idx": 8280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906713.153, "dur": 0.052,
    "args": {
      "Python parent id": 5379, "Python id": 5381, "finished": true, "Ev Idx": 8281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906713.396, "dur": 0.067,
    "args": {
      "Python parent id": 5379, "Python id": 5382, "finished": true, "Ev Idx": 8282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906713.761, "dur": 0.036,
    "args": {
      "Python parent id": 5379, "Python id": 5383, "finished": true, "Ev Idx": 8283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.080, "dur": 4.919,
    "args": {
      "Python parent id": 5379, "Python id": 5384, "finished": true, "Ev Idx": 8284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.283, "dur": 4.660,
    "args": {
      "Python parent id": 5384, "Python id": 5385, "finished": true, "Ev Idx": 8285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.403, "dur": 2.952,
    "args": {
      "Python parent id": 5385, "Python id": 5386, "finished": true, "Ev Idx": 8286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.625, "dur": 2.688,
    "args": {
      "Python parent id": 5386, "Python id": 5387, "finished": true, "Ev Idx": 8287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.744, "dur": 0.247,
    "args": {
      "Python parent id": 5387, "Python id": 5388, "finished": true, "Ev Idx": 8288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906714.889, "dur": 0.054,
    "args": {
      "Python parent id": 5388, "Python id": 5389, "finished": true, "Ev Idx": 8289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.164, "dur": 1.878,
    "args": {
      "Python parent id": 5387, "Python id": 5390, "finished": true, "Ev Idx": 8290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.315, "dur": 1.668,
    "args": {
      "Python parent id": 5390, "Python id": 5391, "finished": true, "Ev Idx": 8291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.547, "dur": 1.381,
    "args": {
      "Python parent id": 5391, "Python id": 5392, "finished": true, "Ev Idx": 8292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.666, "dur": 1.020,
    "args": {
      "Python parent id": 5392, "Python id": 5393, "finished": true, "Ev Idx": 8293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.839, "dur": 0.411,
    "args": {
      "Python parent id": 5393, "Python id": 5394, "finished": true, "Ev Idx": 8294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906715.919, "dur": 0.069,
    "args": {
      "Python parent id": 5394, "Python id": 5395, "finished": true, "Ev Idx": 8295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906716.112, "dur": 0.096,
    "args": {
      "Python parent id": 5394, "Python id": 5396, "finished": true, "Ev Idx": 8296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906717.217, "dur": 0.065,
    "args": {
      "Python parent id": 5387, "Python id": 5397, "finished": true, "Ev Idx": 8297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906717.496, "dur": 0.106,
    "args": {
      "Python parent id": 5385, "Python id": 5398, "finished": true, "Ev Idx": 8298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906717.834, "dur": 1.070,
    "args": {
      "Python parent id": 5385, "Python id": 5399, "finished": true, "Ev Idx": 8299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906717.997, "dur": 0.866,
    "args": {
      "Python parent id": 5399, "Python id": 5400, "finished": true, "Ev Idx": 8300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906718.123, "dur": 0.356,
    "args": {
      "Python parent id": 5400, "Python id": 5401, "finished": true, "Ev Idx": 8301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906718.256, "dur": 0.168,
    "args": {
      "Python parent id": 5401, "Python id": 5402, "finished": true, "Ev Idx": 8302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906718.345, "dur": 0.047,
    "args": {
      "Python parent id": 5402, "Python id": 5403, "finished": true, "Ev Idx": 8303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906718.603, "dur": 0.235,
    "args": {
      "Python parent id": 5400, "Python id": 5404, "finished": true, "Ev Idx": 8304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906719.273, "dur": 0.157,
    "args": {
      "Python parent id": 5369, "Python id": 5405, "finished": true, "Ev Idx": 8305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906719.944, "dur": 0.899,
    "args": {
      "Python parent id": 5369, "Python id": 5406, "finished": true, "Ev Idx": 8306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906720.340, "dur": 0.455,
    "args": {
      "Python parent id": 5406, "Python id": 5407, "finished": true, "Ev Idx": 8307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906721.217, "dur": 2.773,
    "args": {
      "Python parent id": 5368, "Python id": 5408, "finished": true, "Ev Idx": 8308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906724.891, "dur": 106.636,
    "args": {
      "Python parent id": 12, "Python id": 5409, "finished": true, "Ev Idx": 8309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906725.263, "dur": 0.353,
    "args": {
      "Python parent id": 5409, "Python id": 5410, "finished": true, "Ev Idx": 8310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906725.490, "dur": 0.080,
    "args": {
      "Python parent id": 5410, "Python id": 5411, "finished": true, "Ev Idx": 8311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906725.932, "dur": 0.457,
    "args": {
      "Python parent id": 5409, "Python id": 5412, "finished": true, "Ev Idx": 8312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906726.625, "dur": 0.251,
    "args": {
      "Python parent id": 5409, "Python id": 5413, "finished": true, "Ev Idx": 8313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906727.076, "dur": 1.463,
    "args": {
      "Python parent id": 5409, "Python id": 5414, "finished": true, "Ev Idx": 8314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906729.127, "dur": 102.304,
    "args": {
      "Python parent id": 5409, "Python id": 5415, "finished": true, "Ev Idx": 8315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906730.106, "dur": 101.161,
    "args": {
      "Python parent id": 5415, "Python id": 5416, "finished": true, "Ev Idx": 8316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906730.418, "dur": 100.778,
    "args": {
      "Python parent id": 5416, "Python id": 5417, "finished": true, "Ev Idx": 8317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906735.529, "dur": 94.502,
    "args": {
      "Python parent id": 5417, "Python id": 5418, "finished": true, "Ev Idx": 8318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906735.851, "dur": 94.119,
    "args": {
      "Python parent id": 5418, "Python id": 5419, "finished": true, "Ev Idx": 8319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906736.408, "dur": 93.443,
    "args": {
      "Python parent id": 5419, "Python id": 5420, "finished": true, "Ev Idx": 8320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906736.922, "dur": 0.413,
    "args": {
      "Python parent id": 5420, "Python id": 5421, "finished": true, "Ev Idx": 8321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906737.185, "dur": 0.108,
    "args": {
      "Python parent id": 5421, "Python id": 5422, "finished": true, "Ev Idx": 8322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906737.640, "dur": 0.360,
    "args": {
      "Python parent id": 5420, "Python id": 5423, "finished": true, "Ev Idx": 8323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906737.887, "dur": 0.069,
    "args": {
      "Python parent id": 5423, "Python id": 5424, "finished": true, "Ev Idx": 8324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906738.315, "dur": 91.454,
    "args": {
      "Python parent id": 5420, "Python id": 5425, "finished": true, "Ev Idx": 8325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906738.889, "dur": 89.705,
    "args": {
      "Python parent id": 5425, "Python id": 5426, "finished": true, "Ev Idx": 8326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906745.627, "dur": 16.437,
    "args": {
      "Python parent id": 5426, "Python id": 5427, "finished": true, "Ev Idx": 8327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906745.757, "dur": 0.118,
    "args": {
      "Python parent id": 5427, "Python id": 5428, "finished": true, "Ev Idx": 8328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906746.130, "dur": 15.856,
    "args": {
      "Python parent id": 5427, "Python id": 5429, "finished": true, "Ev Idx": 8329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906746.261, "dur": 7.071,
    "args": {
      "Python parent id": 5429, "Python id": 5430, "finished": true, "Ev Idx": 8330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906770.032, "dur": 6.522,
    "args": {
      "Python parent id": 5426, "Python id": 5431, "finished": true, "Ev Idx": 8331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906805.274, "dur": 8.438,
    "args": {
      "Python parent id": 5426, "Python id": 5432, "finished": true, "Ev Idx": 8332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906832.865, "dur": 15.227,
    "args": {
      "Python parent id": 12, "Python id": 5433, "finished": true, "Ev Idx": 8333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906833.183, "dur": 11.459,
    "args": {
      "Python parent id": 5433, "Python id": 5434, "finished": true, "Ev Idx": 8334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906833.385, "dur": 0.496,
    "args": {
      "Python parent id": 5434, "Python id": 5435, "finished": true, "Ev Idx": 8335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906833.565, "dur": 0.223,
    "args": {
      "Python parent id": 5435, "Python id": 5436, "finished": true, "Ev Idx": 8336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906833.667, "dur": 0.071,
    "args": {
      "Python parent id": 5436, "Python id": 5437, "finished": true, "Ev Idx": 8337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906834.233, "dur": 8.467,
    "args": {
      "Python parent id": 5434, "Python id": 5438, "finished": true, "Ev Idx": 8338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906834.401, "dur": 0.059,
    "args": {
      "Python parent id": 5438, "Python id": 5439, "finished": true, "Ev Idx": 8339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906834.562, "dur": 0.056,
    "args": {
      "Python parent id": 5438, "Python id": 5440, "finished": true, "Ev Idx": 8340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906834.749, "dur": 0.056,
    "args": {
      "Python parent id": 5438, "Python id": 5441, "finished": true, "Ev Idx": 8341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906835.285, "dur": 0.074,
    "args": {
      "Python parent id": 5438, "Python id": 5442, "finished": true, "Ev Idx": 8342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906835.605, "dur": 0.072,
    "args": {
      "Python parent id": 5438, "Python id": 5443, "finished": true, "Ev Idx": 8343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906836.003, "dur": 6.641,
    "args": {
      "Python parent id": 5438, "Python id": 5444, "finished": true, "Ev Idx": 8344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906836.098, "dur": 0.061,
    "args": {
      "Python parent id": 5444, "Python id": 5445, "finished": true, "Ev Idx": 8345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906836.308, "dur": 0.051,
    "args": {
      "Python parent id": 5444, "Python id": 5446, "finished": true, "Ev Idx": 8346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906836.537, "dur": 0.066,
    "args": {
      "Python parent id": 5444, "Python id": 5447, "finished": true, "Ev Idx": 8347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906836.919, "dur": 0.028,
    "args": {
      "Python parent id": 5444, "Python id": 5448, "finished": true, "Ev Idx": 8348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906837.296, "dur": 5.246,
    "args": {
      "Python parent id": 5444, "Python id": 5449, "finished": true, "Ev Idx": 8349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906837.564, "dur": 4.922,
    "args": {
      "Python parent id": 5449, "Python id": 5450, "finished": true, "Ev Idx": 8350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906837.678, "dur": 3.149,
    "args": {
      "Python parent id": 5450, "Python id": 5451, "finished": true, "Ev Idx": 8351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906837.848, "dur": 2.936,
    "args": {
      "Python parent id": 5451, "Python id": 5452, "finished": true, "Ev Idx": 8352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906837.999, "dur": 0.236,
    "args": {
      "Python parent id": 5452, "Python id": 5453, "finished": true, "Ev Idx": 8353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906838.124, "dur": 0.061,
    "args": {
      "Python parent id": 5453, "Python id": 5454, "finished": true, "Ev Idx": 8354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906838.417, "dur": 2.044,
    "args": {
      "Python parent id": 5452, "Python id": 5455, "finished": true, "Ev Idx": 8355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906838.595, "dur": 1.797,
    "args": {
      "Python parent id": 5455, "Python id": 5456, "finished": true, "Ev Idx": 8356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906838.849, "dur": 1.482,
    "args": {
      "Python parent id": 5456, "Python id": 5457, "finished": true, "Ev Idx": 8357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906839.014, "dur": 1.089,
    "args": {
      "Python parent id": 5457, "Python id": 5458, "finished": true, "Ev Idx": 8358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906839.205, "dur": 0.464,
    "args": {
      "Python parent id": 5458, "Python id": 5459, "finished": true, "Ev Idx": 8359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906839.289, "dur": 0.069,
    "args": {
      "Python parent id": 5459, "Python id": 5460, "finished": true, "Ev Idx": 8360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906839.491, "dur": 0.138,
    "args": {
      "Python parent id": 5459, "Python id": 5461, "finished": true, "Ev Idx": 8361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906840.649, "dur": 0.105,
    "args": {
      "Python parent id": 5452, "Python id": 5462, "finished": true, "Ev Idx": 8362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906840.968, "dur": 0.120,
    "args": {
      "Python parent id": 5450, "Python id": 5463, "finished": true, "Ev Idx": 8363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906841.341, "dur": 1.103,
    "args": {
      "Python parent id": 5450, "Python id": 5464, "finished": true, "Ev Idx": 8364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906841.496, "dur": 0.905,
    "args": {
      "Python parent id": 5464, "Python id": 5465, "finished": true, "Ev Idx": 8365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906841.624, "dur": 0.385,
    "args": {
      "Python parent id": 5465, "Python id": 5466, "finished": true, "Ev Idx": 8366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906841.794, "dur": 0.160,
    "args": {
      "Python parent id": 5466, "Python id": 5467, "finished": true, "Ev Idx": 8367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906841.871, "dur": 0.054,
    "args": {
      "Python parent id": 5467, "Python id": 5468, "finished": true, "Ev Idx": 8368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906842.128, "dur": 0.245,
    "args": {
      "Python parent id": 5465, "Python id": 5469, "finished": true, "Ev Idx": 8369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906842.819, "dur": 0.184,
    "args": {
      "Python parent id": 5434, "Python id": 5470, "finished": true, "Ev Idx": 8370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906843.522, "dur": 1.006,
    "args": {
      "Python parent id": 5434, "Python id": 5471, "finished": true, "Ev Idx": 8371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906843.956, "dur": 0.502,
    "args": {
      "Python parent id": 5471, "Python id": 5472, "finished": true, "Ev Idx": 8372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906844.906, "dur": 3.132,
    "args": {
      "Python parent id": 5433, "Python id": 5473, "finished": true, "Ev Idx": 8373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906848.628, "dur": 7.204,
    "args": {
      "Python parent id": 12, "Python id": 5474, "finished": true, "Ev Idx": 8374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906856.254, "dur": 0.104,
    "args": {
      "Python parent id": 12, "Python id": 5475, "finished": true, "Ev Idx": 8375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906856.948, "dur": 0.187,
    "args": {
      "Python parent id": 12, "Python id": 5476, "finished": true, "Ev Idx": 8376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906857.642, "dur": 0.067,
    "args": {
      "Python parent id": 12, "Python id": 5477, "finished": true, "Ev Idx": 8377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906858.069, "dur": 8.953,
    "args": {
      "Python parent id": 12, "Python id": 5478, "finished": true, "Ev Idx": 8378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906868.224, "dur": 13.869,
    "args": {
      "Python parent id": 12, "Python id": 5479, "finished": true, "Ev Idx": 8379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906868.501, "dur": 10.438,
    "args": {
      "Python parent id": 5479, "Python id": 5480, "finished": true, "Ev Idx": 8380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906868.662, "dur": 0.451,
    "args": {
      "Python parent id": 5480, "Python id": 5481, "finished": true, "Ev Idx": 8381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906868.842, "dur": 0.209,
    "args": {
      "Python parent id": 5481, "Python id": 5482, "finished": true, "Ev Idx": 8382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906868.928, "dur": 0.088,
    "args": {
      "Python parent id": 5482, "Python id": 5483, "finished": true, "Ev Idx": 8383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906869.403, "dur": 7.755,
    "args": {
      "Python parent id": 5480, "Python id": 5484, "finished": true, "Ev Idx": 8384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906869.481, "dur": 0.104,
    "args": {
      "Python parent id": 5484, "Python id": 5485, "finished": true, "Ev Idx": 8385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906869.695, "dur": 0.059,
    "args": {
      "Python parent id": 5484, "Python id": 5486, "finished": true, "Ev Idx": 8386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906869.884, "dur": 0.057,
    "args": {
      "Python parent id": 5484, "Python id": 5487, "finished": true, "Ev Idx": 8387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906870.310, "dur": 0.082,
    "args": {
      "Python parent id": 5484, "Python id": 5488, "finished": true, "Ev Idx": 8388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906870.612, "dur": 0.069,
    "args": {
      "Python parent id": 5484, "Python id": 5489, "finished": true, "Ev Idx": 8389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906871.077, "dur": 6.016,
    "args": {
      "Python parent id": 5484, "Python id": 5490, "finished": true, "Ev Idx": 8390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906871.180, "dur": 0.059,
    "args": {
      "Python parent id": 5490, "Python id": 5491, "finished": true, "Ev Idx": 8391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906871.378, "dur": 0.056,
    "args": {
      "Python parent id": 5490, "Python id": 5492, "finished": true, "Ev Idx": 8392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906871.586, "dur": 0.064,
    "args": {
      "Python parent id": 5490, "Python id": 5493, "finished": true, "Ev Idx": 8393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906871.952, "dur": 0.032,
    "args": {
      "Python parent id": 5490, "Python id": 5494, "finished": true, "Ev Idx": 8394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906872.289, "dur": 4.751,
    "args": {
      "Python parent id": 5490, "Python id": 5495, "finished": true, "Ev Idx": 8395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906872.478, "dur": 4.506,
    "args": {
      "Python parent id": 5495, "Python id": 5496, "finished": true, "Ev Idx": 8396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906872.591, "dur": 2.721,
    "args": {
      "Python parent id": 5496, "Python id": 5497, "finished": true, "Ev Idx": 8397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906872.760, "dur": 2.507,
    "args": {
      "Python parent id": 5497, "Python id": 5498, "finished": true, "Ev Idx": 8398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906872.904, "dur": 0.218,
    "args": {
      "Python parent id": 5498, "Python id": 5499, "finished": true, "Ev Idx": 8399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.008, "dur": 0.064,
    "args": {
      "Python parent id": 5499, "Python id": 5500, "finished": true, "Ev Idx": 8400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.264, "dur": 1.699,
    "args": {
      "Python parent id": 5498, "Python id": 5501, "finished": true, "Ev Idx": 8401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.453, "dur": 1.456,
    "args": {
      "Python parent id": 5501, "Python id": 5502, "finished": true, "Ev Idx": 8402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.657, "dur": 1.207,
    "args": {
      "Python parent id": 5502, "Python id": 5503, "finished": true, "Ev Idx": 8403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.784, "dur": 0.924,
    "args": {
      "Python parent id": 5503, "Python id": 5504, "finished": true, "Ev Idx": 8404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.924, "dur": 0.395,
    "args": {
      "Python parent id": 5504, "Python id": 5505, "finished": true, "Ev Idx": 8405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906873.998, "dur": 0.042,
    "args": {
      "Python parent id": 5505, "Python id": 5506, "finished": true, "Ev Idx": 8406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906874.162, "dur": 0.119,
    "args": {
      "Python parent id": 5505, "Python id": 5507, "finished": true, "Ev Idx": 8407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906875.144, "dur": 0.091,
    "args": {
      "Python parent id": 5498, "Python id": 5508, "finished": true, "Ev Idx": 8408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906875.477, "dur": 0.066,
    "args": {
      "Python parent id": 5496, "Python id": 5509, "finished": true, "Ev Idx": 8409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906875.794, "dur": 1.146,
    "args": {
      "Python parent id": 5496, "Python id": 5510, "finished": true, "Ev Idx": 8410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906875.949, "dur": 0.947,
    "args": {
      "Python parent id": 5510, "Python id": 5511, "finished": true, "Ev Idx": 8411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906876.115, "dur": 0.356,
    "args": {
      "Python parent id": 5511, "Python id": 5512, "finished": true, "Ev Idx": 8412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906876.245, "dur": 0.168,
    "args": {
      "Python parent id": 5512, "Python id": 5513, "finished": true, "Ev Idx": 8413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906876.331, "dur": 0.053,
    "args": {
      "Python parent id": 5513, "Python id": 5514, "finished": true, "Ev Idx": 8414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906876.587, "dur": 0.278,
    "args": {
      "Python parent id": 5511, "Python id": 5515, "finished": true, "Ev Idx": 8415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906877.272, "dur": 0.152,
    "args": {
      "Python parent id": 5480, "Python id": 5516, "finished": true, "Ev Idx": 8416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906877.942, "dur": 0.880,
    "args": {
      "Python parent id": 5480, "Python id": 5517, "finished": true, "Ev Idx": 8417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906878.306, "dur": 0.445,
    "args": {
      "Python parent id": 5517, "Python id": 5518, "finished": true, "Ev Idx": 8418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906879.174, "dur": 2.866,
    "args": {
      "Python parent id": 5479, "Python id": 5519, "finished": true, "Ev Idx": 8419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740906882.981, "dur": 85.335,
    "args": {
      "Python parent id": 12, "Python id": 5520, "finished": true, "Ev Idx": 8420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906883.311, "dur": 0.429,
    "args": {
      "Python parent id": 5520, "Python id": 5521, "finished": true, "Ev Idx": 8421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906883.612, "dur": 0.084,
    "args": {
      "Python parent id": 5521, "Python id": 5522, "finished": true, "Ev Idx": 8422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906884.032, "dur": 0.389,
    "args": {
      "Python parent id": 5520, "Python id": 5523, "finished": true, "Ev Idx": 8423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740906884.643, "dur": 0.234,
    "args": {
      "Python parent id": 5520, "Python id": 5524, "finished": true, "Ev Idx": 8424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740906885.129, "dur": 1.504,
    "args": {
      "Python parent id": 5520, "Python id": 5525, "finished": true, "Ev Idx": 8425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740906887.270, "dur": 80.940,
    "args": {
      "Python parent id": 5520, "Python id": 5526, "finished": true, "Ev Idx": 8426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740906888.038, "dur": 80.011,
    "args": {
      "Python parent id": 5526, "Python id": 5527, "finished": true, "Ev Idx": 8427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906888.349, "dur": 79.622,
    "args": {
      "Python parent id": 5527, "Python id": 5528, "finished": true, "Ev Idx": 8428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740906893.134, "dur": 73.615,
    "args": {
      "Python parent id": 5528, "Python id": 5529, "finished": true, "Ev Idx": 8429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740906893.423, "dur": 73.279,
    "args": {
      "Python parent id": 5529, "Python id": 5530, "finished": true, "Ev Idx": 8430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740906893.944, "dur": 72.636,
    "args": {
      "Python parent id": 5530, "Python id": 5531, "finished": true, "Ev Idx": 8431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906894.399, "dur": 0.391,
    "args": {
      "Python parent id": 5531, "Python id": 5532, "finished": true, "Ev Idx": 8432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906894.635, "dur": 0.055,
    "args": {
      "Python parent id": 5532, "Python id": 5533, "finished": true, "Ev Idx": 8433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740906895.096, "dur": 0.294,
    "args": {
      "Python parent id": 5531, "Python id": 5534, "finished": true, "Ev Idx": 8434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740906895.293, "dur": 0.046,
    "args": {
      "Python parent id": 5534, "Python id": 5535, "finished": true, "Ev Idx": 8435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906895.729, "dur": 70.736,
    "args": {
      "Python parent id": 5531, "Python id": 5536, "finished": true, "Ev Idx": 8436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740906896.336, "dur": 68.912,
    "args": {
      "Python parent id": 5536, "Python id": 5537, "finished": true, "Ev Idx": 8437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740906903.131, "dur": 16.529,
    "args": {
      "Python parent id": 5537, "Python id": 5538, "finished": true, "Ev Idx": 8438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740906903.256, "dur": 0.133,
    "args": {
      "Python parent id": 5538, "Python id": 5539, "finished": true, "Ev Idx": 8439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740906903.666, "dur": 15.940,
    "args": {
      "Python parent id": 5538, "Python id": 5540, "finished": true, "Ev Idx": 8440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906903.780, "dur": 7.093,
    "args": {
      "Python parent id": 5540, "Python id": 5541, "finished": true, "Ev Idx": 8441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906927.841, "dur": 6.707,
    "args": {
      "Python parent id": 5537, "Python id": 5542, "finished": true, "Ev Idx": 8442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740906942.295, "dur": 8.788,
    "args": {
      "Python parent id": 5537, "Python id": 5543, "finished": true, "Ev Idx": 8443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740906969.412, "dur": 15.441,
    "args": {
      "Python parent id": 12, "Python id": 5544, "finished": true, "Ev Idx": 8444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740906969.822, "dur": 11.625,
    "args": {
      "Python parent id": 5544, "Python id": 5545, "finished": true, "Ev Idx": 8445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906969.995, "dur": 0.544,
    "args": {
      "Python parent id": 5545, "Python id": 5546, "finished": true, "Ev Idx": 8446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906970.204, "dur": 0.255,
    "args": {
      "Python parent id": 5546, "Python id": 5547, "finished": true, "Ev Idx": 8447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906970.332, "dur": 0.091,
    "args": {
      "Python parent id": 5547, "Python id": 5548, "finished": true, "Ev Idx": 8448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906970.854, "dur": 8.611,
    "args": {
      "Python parent id": 5545, "Python id": 5549, "finished": true, "Ev Idx": 8449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906970.990, "dur": 0.068,
    "args": {
      "Python parent id": 5549, "Python id": 5550, "finished": true, "Ev Idx": 8450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906971.145, "dur": 0.059,
    "args": {
      "Python parent id": 5549, "Python id": 5551, "finished": true, "Ev Idx": 8451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906971.342, "dur": 0.059,
    "args": {
      "Python parent id": 5549, "Python id": 5552, "finished": true, "Ev Idx": 8452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906971.880, "dur": 0.062,
    "args": {
      "Python parent id": 5549, "Python id": 5553, "finished": true, "Ev Idx": 8453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906972.193, "dur": 0.076,
    "args": {
      "Python parent id": 5549, "Python id": 5554, "finished": true, "Ev Idx": 8454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906972.642, "dur": 6.772,
    "args": {
      "Python parent id": 5549, "Python id": 5555, "finished": true, "Ev Idx": 8455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906972.729, "dur": 0.058,
    "args": {
      "Python parent id": 5555, "Python id": 5556, "finished": true, "Ev Idx": 8456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906972.963, "dur": 0.052,
    "args": {
      "Python parent id": 5555, "Python id": 5557, "finished": true, "Ev Idx": 8457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906973.190, "dur": 0.061,
    "args": {
      "Python parent id": 5555, "Python id": 5558, "finished": true, "Ev Idx": 8458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740906973.605, "dur": 0.043,
    "args": {
      "Python parent id": 5555, "Python id": 5559, "finished": true, "Ev Idx": 8459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740906973.970, "dur": 5.346,
    "args": {
      "Python parent id": 5555, "Python id": 5560, "finished": true, "Ev Idx": 8460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740906974.279, "dur": 4.982,
    "args": {
      "Python parent id": 5560, "Python id": 5561, "finished": true, "Ev Idx": 8461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740906974.401, "dur": 3.115,
    "args": {
      "Python parent id": 5561, "Python id": 5562, "finished": true, "Ev Idx": 8462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740906974.594, "dur": 2.879,
    "args": {
      "Python parent id": 5562, "Python id": 5563, "finished": true, "Ev Idx": 8463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740906974.721, "dur": 0.232,
    "args": {
      "Python parent id": 5563, "Python id": 5564, "finished": true, "Ev Idx": 8464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740906974.836, "dur": 0.065,
    "args": {
      "Python parent id": 5564, "Python id": 5565, "finished": true, "Ev Idx": 8465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.094, "dur": 2.053,
    "args": {
      "Python parent id": 5563, "Python id": 5566, "finished": true, "Ev Idx": 8466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.284, "dur": 1.788,
    "args": {
      "Python parent id": 5566, "Python id": 5567, "finished": true, "Ev Idx": 8467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.512, "dur": 1.496,
    "args": {
      "Python parent id": 5567, "Python id": 5568, "finished": true, "Ev Idx": 8468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.672, "dur": 1.131,
    "args": {
      "Python parent id": 5568, "Python id": 5569, "finished": true, "Ev Idx": 8469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.859, "dur": 0.472,
    "args": {
      "Python parent id": 5569, "Python id": 5570, "finished": true, "Ev Idx": 8470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906975.942, "dur": 0.049,
    "args": {
      "Python parent id": 5570, "Python id": 5571, "finished": true, "Ev Idx": 8471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740906976.134, "dur": 0.151,
    "args": {
      "Python parent id": 5570, "Python id": 5572, "finished": true, "Ev Idx": 8472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740906977.325, "dur": 0.112,
    "args": {
      "Python parent id": 5563, "Python id": 5573, "finished": true, "Ev Idx": 8473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740906977.666, "dur": 0.097,
    "args": {
      "Python parent id": 5561, "Python id": 5574, "finished": true, "Ev Idx": 8474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740906977.999, "dur": 1.223,
    "args": {
      "Python parent id": 5561, "Python id": 5575, "finished": true, "Ev Idx": 8475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740906978.139, "dur": 1.041,
    "args": {
      "Python parent id": 5575, "Python id": 5576, "finished": true, "Ev Idx": 8476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740906978.284, "dur": 0.402,
    "args": {
      "Python parent id": 5576, "Python id": 5577, "finished": true, "Ev Idx": 8477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740906978.440, "dur": 0.192,
    "args": {
      "Python parent id": 5577, "Python id": 5578, "finished": true, "Ev Idx": 8478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740906978.543, "dur": 0.054,
    "args": {
      "Python parent id": 5578, "Python id": 5579, "finished": true, "Ev Idx": 8479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740906978.804, "dur": 0.348,
    "args": {
      "Python parent id": 5576, "Python id": 5580, "finished": true, "Ev Idx": 8480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740906979.578, "dur": 0.218,
    "args": {
      "Python parent id": 5545, "Python id": 5581, "finished": true, "Ev Idx": 8481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740906980.327, "dur": 0.995,
    "args": {
      "Python parent id": 5545, "Python id": 5582, "finished": true, "Ev Idx": 8482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740906980.785, "dur": 0.488,
    "args": {
      "Python parent id": 5582, "Python id": 5583, "finished": true, "Ev Idx": 8483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740906981.736, "dur": 3.064,
    "args": {
      "Python parent id": 5544, "Python id": 5584, "finished": true, "Ev Idx": 8484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740906985.430, "dur": 7.138,
    "args": {
      "Python parent id": 12, "Python id": 5585, "finished": true, "Ev Idx": 8485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740906992.957, "dur": 0.088,
    "args": {
      "Python parent id": 12, "Python id": 5586, "finished": true, "Ev Idx": 8486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906993.629, "dur": 0.201,
    "args": {
      "Python parent id": 12, "Python id": 5587, "finished": true, "Ev Idx": 8487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740906994.319, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 5588, "finished": true, "Ev Idx": 8488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740906994.771, "dur": 9.163,
    "args": {
      "Python parent id": 12, "Python id": 5589, "finished": true, "Ev Idx": 8489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907005.119, "dur": 14.254,
    "args": {
      "Python parent id": 12, "Python id": 5590, "finished": true, "Ev Idx": 8490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907005.414, "dur": 10.495,
    "args": {
      "Python parent id": 5590, "Python id": 5591, "finished": true, "Ev Idx": 8491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907005.596, "dur": 0.449,
    "args": {
      "Python parent id": 5591, "Python id": 5592, "finished": true, "Ev Idx": 8492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907005.761, "dur": 0.219,
    "args": {
      "Python parent id": 5592, "Python id": 5593, "finished": true, "Ev Idx": 8493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907005.858, "dur": 0.088,
    "args": {
      "Python parent id": 5593, "Python id": 5594, "finished": true, "Ev Idx": 8494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907006.325, "dur": 7.910,
    "args": {
      "Python parent id": 5591, "Python id": 5595, "finished": true, "Ev Idx": 8495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907006.404, "dur": 0.100,
    "args": {
      "Python parent id": 5595, "Python id": 5596, "finished": true, "Ev Idx": 8496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907006.613, "dur": 0.058,
    "args": {
      "Python parent id": 5595, "Python id": 5597, "finished": true, "Ev Idx": 8497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907006.819, "dur": 0.054,
    "args": {
      "Python parent id": 5595, "Python id": 5598, "finished": true, "Ev Idx": 8498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907007.294, "dur": 0.027,
    "args": {
      "Python parent id": 5595, "Python id": 5599, "finished": true, "Ev Idx": 8499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907007.579, "dur": 0.080,
    "args": {
      "Python parent id": 5595, "Python id": 5600, "finished": true, "Ev Idx": 8500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907007.997, "dur": 6.172,
    "args": {
      "Python parent id": 5595, "Python id": 5601, "finished": true, "Ev Idx": 8501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907008.092, "dur": 0.063,
    "args": {
      "Python parent id": 5601, "Python id": 5602, "finished": true, "Ev Idx": 8502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907008.350, "dur": 0.057,
    "args": {
      "Python parent id": 5601, "Python id": 5603, "finished": true, "Ev Idx": 8503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907008.603, "dur": 0.066,
    "args": {
      "Python parent id": 5601, "Python id": 5604, "finished": true, "Ev Idx": 8504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907009.032, "dur": 0.027,
    "args": {
      "Python parent id": 5601, "Python id": 5605, "finished": true, "Ev Idx": 8505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907009.363, "dur": 4.730,
    "args": {
      "Python parent id": 5601, "Python id": 5606, "finished": true, "Ev Idx": 8506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907009.591, "dur": 4.442,
    "args": {
      "Python parent id": 5606, "Python id": 5607, "finished": true, "Ev Idx": 8507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907009.706, "dur": 2.756,
    "args": {
      "Python parent id": 5607, "Python id": 5608, "finished": true, "Ev Idx": 8508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907009.874, "dur": 2.543,
    "args": {
      "Python parent id": 5608, "Python id": 5609, "finished": true, "Ev Idx": 8509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.022, "dur": 0.215,
    "args": {
      "Python parent id": 5609, "Python id": 5610, "finished": true, "Ev Idx": 8510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.124, "dur": 0.068,
    "args": {
      "Python parent id": 5610, "Python id": 5611, "finished": true, "Ev Idx": 8511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.380, "dur": 1.767,
    "args": {
      "Python parent id": 5609, "Python id": 5612, "finished": true, "Ev Idx": 8512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.536, "dur": 1.555,
    "args": {
      "Python parent id": 5612, "Python id": 5613, "finished": true, "Ev Idx": 8513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.735, "dur": 1.310,
    "args": {
      "Python parent id": 5613, "Python id": 5614, "finished": true, "Ev Idx": 8514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.846, "dur": 1.002,
    "args": {
      "Python parent id": 5614, "Python id": 5615, "finished": true, "Ev Idx": 8515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907010.999, "dur": 0.437,
    "args": {
      "Python parent id": 5615, "Python id": 5616, "finished": true, "Ev Idx": 8516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907011.080, "dur": 0.064,
    "args": {
      "Python parent id": 5616, "Python id": 5617, "finished": true, "Ev Idx": 8517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907011.294, "dur": 0.102,
    "args": {
      "Python parent id": 5616, "Python id": 5618, "finished": true, "Ev Idx": 8518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907012.317, "dur": 0.066,
    "args": {
      "Python parent id": 5609, "Python id": 5619, "finished": true, "Ev Idx": 8519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907012.623, "dur": 0.086,
    "args": {
      "Python parent id": 5607, "Python id": 5620, "finished": true, "Ev Idx": 8520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907012.941, "dur": 1.051,
    "args": {
      "Python parent id": 5607, "Python id": 5621, "finished": true, "Ev Idx": 8521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907013.066, "dur": 0.884,
    "args": {
      "Python parent id": 5621, "Python id": 5622, "finished": true, "Ev Idx": 8522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907013.197, "dur": 0.354,
    "args": {
      "Python parent id": 5622, "Python id": 5623, "finished": true, "Ev Idx": 8523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907013.324, "dur": 0.171,
    "args": {
      "Python parent id": 5623, "Python id": 5624, "finished": true, "Ev Idx": 8524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907013.407, "dur": 0.055,
    "args": {
      "Python parent id": 5624, "Python id": 5625, "finished": true, "Ev Idx": 8525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907013.683, "dur": 0.236,
    "args": {
      "Python parent id": 5622, "Python id": 5626, "finished": true, "Ev Idx": 8526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907014.347, "dur": 0.130,
    "args": {
      "Python parent id": 5591, "Python id": 5627, "finished": true, "Ev Idx": 8527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907014.974, "dur": 0.820,
    "args": {
      "Python parent id": 5591, "Python id": 5628, "finished": true, "Ev Idx": 8528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907015.350, "dur": 0.392,
    "args": {
      "Python parent id": 5628, "Python id": 5629, "finished": true, "Ev Idx": 8529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907016.186, "dur": 3.134,
    "args": {
      "Python parent id": 5590, "Python id": 5630, "finished": true, "Ev Idx": 8530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907020.327, "dur": 106.446,
    "args": {
      "Python parent id": 12, "Python id": 5631, "finished": true, "Ev Idx": 8531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907020.665, "dur": 0.428,
    "args": {
      "Python parent id": 5631, "Python id": 5632, "finished": true, "Ev Idx": 8532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907020.947, "dur": 0.100,
    "args": {
      "Python parent id": 5632, "Python id": 5633, "finished": true, "Ev Idx": 8533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907021.376, "dur": 0.325,
    "args": {
      "Python parent id": 5631, "Python id": 5634, "finished": true, "Ev Idx": 8534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907021.876, "dur": 0.271,
    "args": {
      "Python parent id": 5631, "Python id": 5635, "finished": true, "Ev Idx": 8535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907022.392, "dur": 2.549,
    "args": {
      "Python parent id": 5631, "Python id": 5636, "finished": true, "Ev Idx": 8536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907025.459, "dur": 101.177,
    "args": {
      "Python parent id": 5631, "Python id": 5637, "finished": true, "Ev Idx": 8537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907026.134, "dur": 100.308,
    "args": {
      "Python parent id": 5637, "Python id": 5638, "finished": true, "Ev Idx": 8538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907026.394, "dur": 99.975,
    "args": {
      "Python parent id": 5638, "Python id": 5639, "finished": true, "Ev Idx": 8539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907032.155, "dur": 92.691,
    "args": {
      "Python parent id": 5639, "Python id": 5640, "finished": true, "Ev Idx": 8540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907032.449, "dur": 92.315,
    "args": {
      "Python parent id": 5640, "Python id": 5641, "finished": true, "Ev Idx": 8541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907032.930, "dur": 91.699,
    "args": {
      "Python parent id": 5641, "Python id": 5642, "finished": true, "Ev Idx": 8542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907033.422, "dur": 0.656,
    "args": {
      "Python parent id": 5642, "Python id": 5643, "finished": true, "Ev Idx": 8543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907033.663, "dur": 0.337,
    "args": {
      "Python parent id": 5643, "Python id": 5644, "finished": true, "Ev Idx": 8544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907034.337, "dur": 0.295,
    "args": {
      "Python parent id": 5642, "Python id": 5645, "finished": true, "Ev Idx": 8545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907034.529, "dur": 0.061,
    "args": {
      "Python parent id": 5645, "Python id": 5646, "finished": true, "Ev Idx": 8546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907034.944, "dur": 89.573,
    "args": {
      "Python parent id": 5642, "Python id": 5647, "finished": true, "Ev Idx": 8547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907035.530, "dur": 87.842,
    "args": {
      "Python parent id": 5647, "Python id": 5648, "finished": true, "Ev Idx": 8548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907042.349, "dur": 18.454,
    "args": {
      "Python parent id": 5648, "Python id": 5649, "finished": true, "Ev Idx": 8549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907042.444, "dur": 1.344,
    "args": {
      "Python parent id": 5649, "Python id": 5650, "finished": true, "Ev Idx": 8550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907044.055, "dur": 16.694,
    "args": {
      "Python parent id": 5649, "Python id": 5651, "finished": true, "Ev Idx": 8551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907044.181, "dur": 7.458,
    "args": {
      "Python parent id": 5651, "Python id": 5652, "finished": true, "Ev Idx": 8552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907069.457, "dur": 7.443,
    "args": {
      "Python parent id": 5648, "Python id": 5653, "finished": true, "Ev Idx": 8553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907098.617, "dur": 9.831,
    "args": {
      "Python parent id": 5648, "Python id": 5654, "finished": true, "Ev Idx": 8554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907127.931, "dur": 15.730,
    "args": {
      "Python parent id": 12, "Python id": 5655, "finished": true, "Ev Idx": 8555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907128.377, "dur": 11.637,
    "args": {
      "Python parent id": 5655, "Python id": 5656, "finished": true, "Ev Idx": 8556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907128.643, "dur": 0.460,
    "args": {
      "Python parent id": 5656, "Python id": 5657, "finished": true, "Ev Idx": 8557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907128.829, "dur": 0.211,
    "args": {
      "Python parent id": 5657, "Python id": 5658, "finished": true, "Ev Idx": 8558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907128.939, "dur": 0.065,
    "args": {
      "Python parent id": 5658, "Python id": 5659, "finished": true, "Ev Idx": 8559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907129.408, "dur": 8.590,
    "args": {
      "Python parent id": 5656, "Python id": 5660, "finished": true, "Ev Idx": 8560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907129.528, "dur": 0.068,
    "args": {
      "Python parent id": 5660, "Python id": 5661, "finished": true, "Ev Idx": 8561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907129.700, "dur": 0.062,
    "args": {
      "Python parent id": 5660, "Python id": 5662, "finished": true, "Ev Idx": 8562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907129.916, "dur": 0.061,
    "args": {
      "Python parent id": 5660, "Python id": 5663, "finished": true, "Ev Idx": 8563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907130.429, "dur": 0.054,
    "args": {
      "Python parent id": 5660, "Python id": 5664, "finished": true, "Ev Idx": 8564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907130.738, "dur": 0.074,
    "args": {
      "Python parent id": 5660, "Python id": 5665, "finished": true, "Ev Idx": 8565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907131.193, "dur": 6.741,
    "args": {
      "Python parent id": 5660, "Python id": 5666, "finished": true, "Ev Idx": 8566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907131.303, "dur": 0.061,
    "args": {
      "Python parent id": 5666, "Python id": 5667, "finished": true, "Ev Idx": 8567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907131.572, "dur": 0.053,
    "args": {
      "Python parent id": 5666, "Python id": 5668, "finished": true, "Ev Idx": 8568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907131.809, "dur": 0.072,
    "args": {
      "Python parent id": 5666, "Python id": 5669, "finished": true, "Ev Idx": 8569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907132.232, "dur": 0.034,
    "args": {
      "Python parent id": 5666, "Python id": 5670, "finished": true, "Ev Idx": 8570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907132.617, "dur": 5.263,
    "args": {
      "Python parent id": 5666, "Python id": 5671, "finished": true, "Ev Idx": 8571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907132.914, "dur": 4.910,
    "args": {
      "Python parent id": 5671, "Python id": 5672, "finished": true, "Ev Idx": 8572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.052, "dur": 3.185,
    "args": {
      "Python parent id": 5672, "Python id": 5673, "finished": true, "Ev Idx": 8573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.250, "dur": 2.949,
    "args": {
      "Python parent id": 5673, "Python id": 5674, "finished": true, "Ev Idx": 8574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.386, "dur": 0.271,
    "args": {
      "Python parent id": 5674, "Python id": 5675, "finished": true, "Ev Idx": 8575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.532, "dur": 0.064,
    "args": {
      "Python parent id": 5675, "Python id": 5676, "finished": true, "Ev Idx": 8576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.788, "dur": 2.063,
    "args": {
      "Python parent id": 5674, "Python id": 5677, "finished": true, "Ev Idx": 8577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907133.993, "dur": 1.788,
    "args": {
      "Python parent id": 5677, "Python id": 5678, "finished": true, "Ev Idx": 8578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907134.183, "dur": 1.531,
    "args": {
      "Python parent id": 5678, "Python id": 5679, "finished": true, "Ev Idx": 8579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907134.441, "dur": 1.064,
    "args": {
      "Python parent id": 5679, "Python id": 5680, "finished": true, "Ev Idx": 8580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907134.614, "dur": 0.434,
    "args": {
      "Python parent id": 5680, "Python id": 5681, "finished": true, "Ev Idx": 8581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907134.694, "dur": 0.051,
    "args": {
      "Python parent id": 5681, "Python id": 5682, "finished": true, "Ev Idx": 8582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907134.883, "dur": 0.124,
    "args": {
      "Python parent id": 5681, "Python id": 5683, "finished": true, "Ev Idx": 8583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907136.034, "dur": 0.115,
    "args": {
      "Python parent id": 5674, "Python id": 5684, "finished": true, "Ev Idx": 8584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907136.407, "dur": 0.071,
    "args": {
      "Python parent id": 5672, "Python id": 5685, "finished": true, "Ev Idx": 8585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907136.704, "dur": 1.079,
    "args": {
      "Python parent id": 5672, "Python id": 5686, "finished": true, "Ev Idx": 8586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907136.845, "dur": 0.898,
    "args": {
      "Python parent id": 5686, "Python id": 5687, "finished": true, "Ev Idx": 8587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907137.017, "dur": 0.345,
    "args": {
      "Python parent id": 5687, "Python id": 5688, "finished": true, "Ev Idx": 8588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907137.134, "dur": 0.174,
    "args": {
      "Python parent id": 5688, "Python id": 5689, "finished": true, "Ev Idx": 8589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907137.218, "dur": 0.057,
    "args": {
      "Python parent id": 5689, "Python id": 5690, "finished": true, "Ev Idx": 8590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907137.477, "dur": 0.237,
    "args": {
      "Python parent id": 5687, "Python id": 5691, "finished": true, "Ev Idx": 8591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907138.123, "dur": 0.232,
    "args": {
      "Python parent id": 5656, "Python id": 5692, "finished": true, "Ev Idx": 8592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907138.916, "dur": 0.958,
    "args": {
      "Python parent id": 5656, "Python id": 5693, "finished": true, "Ev Idx": 8593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907139.371, "dur": 0.453,
    "args": {
      "Python parent id": 5693, "Python id": 5694, "finished": true, "Ev Idx": 8594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907140.306, "dur": 3.285,
    "args": {
      "Python parent id": 5655, "Python id": 5695, "finished": true, "Ev Idx": 8595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907144.238, "dur": 8.066,
    "args": {
      "Python parent id": 12, "Python id": 5696, "finished": true, "Ev Idx": 8596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907152.684, "dur": 0.107,
    "args": {
      "Python parent id": 12, "Python id": 5697, "finished": true, "Ev Idx": 8597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907153.327, "dur": 0.185,
    "args": {
      "Python parent id": 12, "Python id": 5698, "finished": true, "Ev Idx": 8598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907153.994, "dur": 0.070,
    "args": {
      "Python parent id": 12, "Python id": 5699, "finished": true, "Ev Idx": 8599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907154.429, "dur": 9.821,
    "args": {
      "Python parent id": 12, "Python id": 5700, "finished": true, "Ev Idx": 8600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907165.675, "dur": 14.325,
    "args": {
      "Python parent id": 12, "Python id": 5701, "finished": true, "Ev Idx": 8601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.032, "dur": 10.465,
    "args": {
      "Python parent id": 5701, "Python id": 5702, "finished": true, "Ev Idx": 8602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.204, "dur": 0.427,
    "args": {
      "Python parent id": 5702, "Python id": 5703, "finished": true, "Ev Idx": 8603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.356, "dur": 0.211,
    "args": {
      "Python parent id": 5703, "Python id": 5704, "finished": true, "Ev Idx": 8604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.446, "dur": 0.088,
    "args": {
      "Python parent id": 5704, "Python id": 5705, "finished": true, "Ev Idx": 8605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.908, "dur": 7.875,
    "args": {
      "Python parent id": 5702, "Python id": 5706, "finished": true, "Ev Idx": 8606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907166.991, "dur": 0.119,
    "args": {
      "Python parent id": 5706, "Python id": 5707, "finished": true, "Ev Idx": 8607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907167.211, "dur": 0.051,
    "args": {
      "Python parent id": 5706, "Python id": 5708, "finished": true, "Ev Idx": 8608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907167.402, "dur": 0.054,
    "args": {
      "Python parent id": 5706, "Python id": 5709, "finished": true, "Ev Idx": 8609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907167.917, "dur": 0.053,
    "args": {
      "Python parent id": 5706, "Python id": 5710, "finished": true, "Ev Idx": 8610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907168.250, "dur": 0.072,
    "args": {
      "Python parent id": 5706, "Python id": 5711, "finished": true, "Ev Idx": 8611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907168.638, "dur": 6.084,
    "args": {
      "Python parent id": 5706, "Python id": 5712, "finished": true, "Ev Idx": 8612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907168.727, "dur": 0.059,
    "args": {
      "Python parent id": 5712, "Python id": 5713, "finished": true, "Ev Idx": 8613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907168.927, "dur": 0.057,
    "args": {
      "Python parent id": 5712, "Python id": 5714, "finished": true, "Ev Idx": 8614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907169.172, "dur": 0.065,
    "args": {
      "Python parent id": 5712, "Python id": 5715, "finished": true, "Ev Idx": 8615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907169.555, "dur": 0.068,
    "args": {
      "Python parent id": 5712, "Python id": 5716, "finished": true, "Ev Idx": 8616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907169.900, "dur": 4.743,
    "args": {
      "Python parent id": 5712, "Python id": 5717, "finished": true, "Ev Idx": 8617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.121, "dur": 4.467,
    "args": {
      "Python parent id": 5717, "Python id": 5718, "finished": true, "Ev Idx": 8618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.246, "dur": 2.782,
    "args": {
      "Python parent id": 5718, "Python id": 5719, "finished": true, "Ev Idx": 8619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.418, "dur": 2.570,
    "args": {
      "Python parent id": 5719, "Python id": 5720, "finished": true, "Ev Idx": 8620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.536, "dur": 0.265,
    "args": {
      "Python parent id": 5720, "Python id": 5721, "finished": true, "Ev Idx": 8621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.663, "dur": 0.063,
    "args": {
      "Python parent id": 5721, "Python id": 5722, "finished": true, "Ev Idx": 8622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907170.937, "dur": 1.773,
    "args": {
      "Python parent id": 5720, "Python id": 5723, "finished": true, "Ev Idx": 8623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.105, "dur": 1.548,
    "args": {
      "Python parent id": 5723, "Python id": 5724, "finished": true, "Ev Idx": 8624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.336, "dur": 1.276,
    "args": {
      "Python parent id": 5724, "Python id": 5725, "finished": true, "Ev Idx": 8625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.469, "dur": 0.971,
    "args": {
      "Python parent id": 5725, "Python id": 5726, "finished": true, "Ev Idx": 8626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.648, "dur": 0.426,
    "args": {
      "Python parent id": 5726, "Python id": 5727, "finished": true, "Ev Idx": 8627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.752, "dur": 0.041,
    "args": {
      "Python parent id": 5727, "Python id": 5728, "finished": true, "Ev Idx": 8628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907171.919, "dur": 0.117,
    "args": {
      "Python parent id": 5727, "Python id": 5729, "finished": true, "Ev Idx": 8629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907172.878, "dur": 0.062,
    "args": {
      "Python parent id": 5720, "Python id": 5730, "finished": true, "Ev Idx": 8630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.181, "dur": 0.082,
    "args": {
      "Python parent id": 5718, "Python id": 5731, "finished": true, "Ev Idx": 8631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.499, "dur": 1.047,
    "args": {
      "Python parent id": 5718, "Python id": 5732, "finished": true, "Ev Idx": 8632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.622, "dur": 0.885,
    "args": {
      "Python parent id": 5732, "Python id": 5733, "finished": true, "Ev Idx": 8633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.754, "dur": 0.367,
    "args": {
      "Python parent id": 5733, "Python id": 5734, "finished": true, "Ev Idx": 8634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.882, "dur": 0.185,
    "args": {
      "Python parent id": 5734, "Python id": 5735, "finished": true, "Ev Idx": 8635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907173.982, "dur": 0.051,
    "args": {
      "Python parent id": 5735, "Python id": 5736, "finished": true, "Ev Idx": 8636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907174.246, "dur": 0.235,
    "args": {
      "Python parent id": 5733, "Python id": 5737, "finished": true, "Ev Idx": 8637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907174.893, "dur": 0.136,
    "args": {
      "Python parent id": 5702, "Python id": 5738, "finished": true, "Ev Idx": 8638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907175.564, "dur": 0.821,
    "args": {
      "Python parent id": 5702, "Python id": 5739, "finished": true, "Ev Idx": 8639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907175.942, "dur": 0.395,
    "args": {
      "Python parent id": 5739, "Python id": 5740, "finished": true, "Ev Idx": 8640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907176.758, "dur": 3.185,
    "args": {
      "Python parent id": 5701, "Python id": 5741, "finished": true, "Ev Idx": 8641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907180.859, "dur": 86.197,
    "args": {
      "Python parent id": 12, "Python id": 5742, "finished": true, "Ev Idx": 8642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907181.168, "dur": 0.527,
    "args": {
      "Python parent id": 5742, "Python id": 5743, "finished": true, "Ev Idx": 8643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907181.445, "dur": 0.207,
    "args": {
      "Python parent id": 5743, "Python id": 5744, "finished": true, "Ev Idx": 8644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907182.055, "dur": 0.343,
    "args": {
      "Python parent id": 5742, "Python id": 5745, "finished": true, "Ev Idx": 8645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907182.622, "dur": 0.274,
    "args": {
      "Python parent id": 5742, "Python id": 5746, "finished": true, "Ev Idx": 8646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907183.124, "dur": 1.448,
    "args": {
      "Python parent id": 5742, "Python id": 5747, "finished": true, "Ev Idx": 8647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907185.115, "dur": 81.847,
    "args": {
      "Python parent id": 5742, "Python id": 5748, "finished": true, "Ev Idx": 8648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907185.862, "dur": 80.747,
    "args": {
      "Python parent id": 5748, "Python id": 5749, "finished": true, "Ev Idx": 8649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907186.136, "dur": 80.425,
    "args": {
      "Python parent id": 5749, "Python id": 5750, "finished": true, "Ev Idx": 8650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907191.155, "dur": 74.143,
    "args": {
      "Python parent id": 5750, "Python id": 5751, "finished": true, "Ev Idx": 8651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907191.452, "dur": 73.752,
    "args": {
      "Python parent id": 5751, "Python id": 5752, "finished": true, "Ev Idx": 8652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907191.957, "dur": 73.083,
    "args": {
      "Python parent id": 5752, "Python id": 5753, "finished": true, "Ev Idx": 8653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907192.482, "dur": 0.422,
    "args": {
      "Python parent id": 5753, "Python id": 5754, "finished": true, "Ev Idx": 8654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907192.775, "dur": 0.068,
    "args": {
      "Python parent id": 5754, "Python id": 5755, "finished": true, "Ev Idx": 8655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907193.176, "dur": 0.340,
    "args": {
      "Python parent id": 5753, "Python id": 5756, "finished": true, "Ev Idx": 8656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907193.414, "dur": 0.061,
    "args": {
      "Python parent id": 5756, "Python id": 5757, "finished": true, "Ev Idx": 8657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907193.831, "dur": 71.109,
    "args": {
      "Python parent id": 5753, "Python id": 5758, "finished": true, "Ev Idx": 8658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907194.436, "dur": 69.376,
    "args": {
      "Python parent id": 5758, "Python id": 5759, "finished": true, "Ev Idx": 8659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907201.478, "dur": 17.064,
    "args": {
      "Python parent id": 5759, "Python id": 5760, "finished": true, "Ev Idx": 8660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907201.608, "dur": 0.188,
    "args": {
      "Python parent id": 5760, "Python id": 5761, "finished": true, "Ev Idx": 8661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907202.131, "dur": 16.319,
    "args": {
      "Python parent id": 5760, "Python id": 5762, "finished": true, "Ev Idx": 8662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907202.250, "dur": 7.421,
    "args": {
      "Python parent id": 5762, "Python id": 5763, "finished": true, "Ev Idx": 8663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907226.539, "dur": 6.738,
    "args": {
      "Python parent id": 5759, "Python id": 5764, "finished": true, "Ev Idx": 8664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907241.458, "dur": 8.488,
    "args": {
      "Python parent id": 5759, "Python id": 5765, "finished": true, "Ev Idx": 8665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907268.216, "dur": 16.111,
    "args": {
      "Python parent id": 12, "Python id": 5766, "finished": true, "Ev Idx": 8666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907268.622, "dur": 12.161,
    "args": {
      "Python parent id": 5766, "Python id": 5767, "finished": true, "Ev Idx": 8667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907268.839, "dur": 0.422,
    "args": {
      "Python parent id": 5767, "Python id": 5768, "finished": true, "Ev Idx": 8668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907268.978, "dur": 0.220,
    "args": {
      "Python parent id": 5768, "Python id": 5769, "finished": true, "Ev Idx": 8669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907269.092, "dur": 0.076,
    "args": {
      "Python parent id": 5769, "Python id": 5770, "finished": true, "Ev Idx": 8670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907269.574, "dur": 9.086,
    "args": {
      "Python parent id": 5767, "Python id": 5771, "finished": true, "Ev Idx": 8671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907269.652, "dur": 0.061,
    "args": {
      "Python parent id": 5771, "Python id": 5772, "finished": true, "Ev Idx": 8672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907269.817, "dur": 0.056,
    "args": {
      "Python parent id": 5771, "Python id": 5773, "finished": true, "Ev Idx": 8673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907270.008, "dur": 0.060,
    "args": {
      "Python parent id": 5771, "Python id": 5774, "finished": true, "Ev Idx": 8674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907270.513, "dur": 0.036,
    "args": {
      "Python parent id": 5771, "Python id": 5775, "finished": true, "Ev Idx": 8675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907270.815, "dur": 0.065,
    "args": {
      "Python parent id": 5771, "Python id": 5776, "finished": true, "Ev Idx": 8676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907271.315, "dur": 7.290,
    "args": {
      "Python parent id": 5771, "Python id": 5777, "finished": true, "Ev Idx": 8677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907271.408, "dur": 0.063,
    "args": {
      "Python parent id": 5777, "Python id": 5778, "finished": true, "Ev Idx": 8678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907271.609, "dur": 0.056,
    "args": {
      "Python parent id": 5777, "Python id": 5779, "finished": true, "Ev Idx": 8679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907271.846, "dur": 0.068,
    "args": {
      "Python parent id": 5777, "Python id": 5780, "finished": true, "Ev Idx": 8680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907272.243, "dur": 0.053,
    "args": {
      "Python parent id": 5777, "Python id": 5781, "finished": true, "Ev Idx": 8681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907272.565, "dur": 5.963,
    "args": {
      "Python parent id": 5777, "Python id": 5782, "finished": true, "Ev Idx": 8682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907272.841, "dur": 5.632,
    "args": {
      "Python parent id": 5782, "Python id": 5783, "finished": true, "Ev Idx": 8683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907273.004, "dur": 3.593,
    "args": {
      "Python parent id": 5783, "Python id": 5784, "finished": true, "Ev Idx": 8684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907273.204, "dur": 3.352,
    "args": {
      "Python parent id": 5784, "Python id": 5785, "finished": true, "Ev Idx": 8685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907273.330, "dur": 0.475,
    "args": {
      "Python parent id": 5785, "Python id": 5786, "finished": true, "Ev Idx": 8686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907273.478, "dur": 0.112,
    "args": {
      "Python parent id": 5786, "Python id": 5787, "finished": true, "Ev Idx": 8687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907273.971, "dur": 2.204,
    "args": {
      "Python parent id": 5785, "Python id": 5788, "finished": true, "Ev Idx": 8688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907274.179, "dur": 1.906,
    "args": {
      "Python parent id": 5788, "Python id": 5789, "finished": true, "Ev Idx": 8689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907274.380, "dur": 1.659,
    "args": {
      "Python parent id": 5789, "Python id": 5790, "finished": true, "Ev Idx": 8690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907274.563, "dur": 1.256,
    "args": {
      "Python parent id": 5790, "Python id": 5791, "finished": true, "Ev Idx": 8691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907274.795, "dur": 0.460,
    "args": {
      "Python parent id": 5791, "Python id": 5792, "finished": true, "Ev Idx": 8692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907274.892, "dur": 0.046,
    "args": {
      "Python parent id": 5792, "Python id": 5793, "finished": true, "Ev Idx": 8693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907275.080, "dur": 0.136,
    "args": {
      "Python parent id": 5792, "Python id": 5794, "finished": true, "Ev Idx": 8694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907276.378, "dur": 0.130,
    "args": {
      "Python parent id": 5785, "Python id": 5795, "finished": true, "Ev Idx": 8695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907276.802, "dur": 0.101,
    "args": {
      "Python parent id": 5783, "Python id": 5796, "finished": true, "Ev Idx": 8696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907277.189, "dur": 1.244,
    "args": {
      "Python parent id": 5783, "Python id": 5797, "finished": true, "Ev Idx": 8697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907277.342, "dur": 1.049,
    "args": {
      "Python parent id": 5797, "Python id": 5798, "finished": true, "Ev Idx": 8698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907277.466, "dur": 0.466,
    "args": {
      "Python parent id": 5798, "Python id": 5799, "finished": true, "Ev Idx": 8699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907277.706, "dur": 0.171,
    "args": {
      "Python parent id": 5799, "Python id": 5800, "finished": true, "Ev Idx": 8700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907277.796, "dur": 0.051,
    "args": {
      "Python parent id": 5800, "Python id": 5801, "finished": true, "Ev Idx": 8701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907278.054, "dur": 0.304,
    "args": {
      "Python parent id": 5798, "Python id": 5802, "finished": true, "Ev Idx": 8702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907278.805, "dur": 0.226,
    "args": {
      "Python parent id": 5767, "Python id": 5803, "finished": true, "Ev Idx": 8703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907279.607, "dur": 0.999,
    "args": {
      "Python parent id": 5767, "Python id": 5804, "finished": true, "Ev Idx": 8704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907280.045, "dur": 0.512,
    "args": {
      "Python parent id": 5804, "Python id": 5805, "finished": true, "Ev Idx": 8705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907281.065, "dur": 3.169,
    "args": {
      "Python parent id": 5766, "Python id": 5806, "finished": true, "Ev Idx": 8706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907284.893, "dur": 7.131,
    "args": {
      "Python parent id": 12, "Python id": 5807, "finished": true, "Ev Idx": 8707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907292.491, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 5808, "finished": true, "Ev Idx": 8708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907293.295, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 5809, "finished": true, "Ev Idx": 8709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907293.944, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 5810, "finished": true, "Ev Idx": 8710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907294.374, "dur": 9.077,
    "args": {
      "Python parent id": 12, "Python id": 5811, "finished": true, "Ev Idx": 8711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907304.629, "dur": 13.920,
    "args": {
      "Python parent id": 12, "Python id": 5812, "finished": true, "Ev Idx": 8712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907304.886, "dur": 10.768,
    "args": {
      "Python parent id": 5812, "Python id": 5813, "finished": true, "Ev Idx": 8713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907305.054, "dur": 0.441,
    "args": {
      "Python parent id": 5813, "Python id": 5814, "finished": true, "Ev Idx": 8714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907305.213, "dur": 0.215,
    "args": {
      "Python parent id": 5814, "Python id": 5815, "finished": true, "Ev Idx": 8715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907305.293, "dur": 0.101,
    "args": {
      "Python parent id": 5815, "Python id": 5816, "finished": true, "Ev Idx": 8716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907305.745, "dur": 8.062,
    "args": {
      "Python parent id": 5813, "Python id": 5817, "finished": true, "Ev Idx": 8717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907305.841, "dur": 0.068,
    "args": {
      "Python parent id": 5817, "Python id": 5818, "finished": true, "Ev Idx": 8718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907306.026, "dur": 0.054,
    "args": {
      "Python parent id": 5817, "Python id": 5819, "finished": true, "Ev Idx": 8719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907306.214, "dur": 0.058,
    "args": {
      "Python parent id": 5817, "Python id": 5820, "finished": true, "Ev Idx": 8720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907306.684, "dur": 0.042,
    "args": {
      "Python parent id": 5817, "Python id": 5821, "finished": true, "Ev Idx": 8721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907306.998, "dur": 0.077,
    "args": {
      "Python parent id": 5817, "Python id": 5822, "finished": true, "Ev Idx": 8722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907307.394, "dur": 6.351,
    "args": {
      "Python parent id": 5817, "Python id": 5823, "finished": true, "Ev Idx": 8723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907307.490, "dur": 0.058,
    "args": {
      "Python parent id": 5823, "Python id": 5824, "finished": true, "Ev Idx": 8724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907307.685, "dur": 0.061,
    "args": {
      "Python parent id": 5823, "Python id": 5825, "finished": true, "Ev Idx": 8725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907307.910, "dur": 0.070,
    "args": {
      "Python parent id": 5823, "Python id": 5826, "finished": true, "Ev Idx": 8726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907308.374, "dur": 0.028,
    "args": {
      "Python parent id": 5823, "Python id": 5827, "finished": true, "Ev Idx": 8727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907308.716, "dur": 4.977,
    "args": {
      "Python parent id": 5823, "Python id": 5828, "finished": true, "Ev Idx": 8728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907308.959, "dur": 4.683,
    "args": {
      "Python parent id": 5828, "Python id": 5829, "finished": true, "Ev Idx": 8729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.121, "dur": 2.885,
    "args": {
      "Python parent id": 5829, "Python id": 5830, "finished": true, "Ev Idx": 8730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.287, "dur": 2.676,
    "args": {
      "Python parent id": 5830, "Python id": 5831, "finished": true, "Ev Idx": 8731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.411, "dur": 0.232,
    "args": {
      "Python parent id": 5831, "Python id": 5832, "finished": true, "Ev Idx": 8732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.523, "dur": 0.061,
    "args": {
      "Python parent id": 5832, "Python id": 5833, "finished": true, "Ev Idx": 8733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.779, "dur": 1.930,
    "args": {
      "Python parent id": 5831, "Python id": 5834, "finished": true, "Ev Idx": 8734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907309.969, "dur": 1.661,
    "args": {
      "Python parent id": 5834, "Python id": 5835, "finished": true, "Ev Idx": 8735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907310.156, "dur": 1.428,
    "args": {
      "Python parent id": 5835, "Python id": 5836, "finished": true, "Ev Idx": 8736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907310.302, "dur": 1.018,
    "args": {
      "Python parent id": 5836, "Python id": 5837, "finished": true, "Ev Idx": 8737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907310.489, "dur": 0.446,
    "args": {
      "Python parent id": 5837, "Python id": 5838, "finished": true, "Ev Idx": 8738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907310.572, "dur": 0.046,
    "args": {
      "Python parent id": 5838, "Python id": 5839, "finished": true, "Ev Idx": 8739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907310.744, "dur": 0.148,
    "args": {
      "Python parent id": 5838, "Python id": 5840, "finished": true, "Ev Idx": 8740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907311.871, "dur": 0.063,
    "args": {
      "Python parent id": 5831, "Python id": 5841, "finished": true, "Ev Idx": 8741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.143, "dur": 0.087,
    "args": {
      "Python parent id": 5829, "Python id": 5842, "finished": true, "Ev Idx": 8742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.483, "dur": 1.093,
    "args": {
      "Python parent id": 5829, "Python id": 5843, "finished": true, "Ev Idx": 8743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.605, "dur": 0.910,
    "args": {
      "Python parent id": 5843, "Python id": 5844, "finished": true, "Ev Idx": 8744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.740, "dur": 0.348,
    "args": {
      "Python parent id": 5844, "Python id": 5845, "finished": true, "Ev Idx": 8745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.865, "dur": 0.169,
    "args": {
      "Python parent id": 5845, "Python id": 5846, "finished": true, "Ev Idx": 8746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907312.948, "dur": 0.057,
    "args": {
      "Python parent id": 5846, "Python id": 5847, "finished": true, "Ev Idx": 8747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907313.203, "dur": 0.280,
    "args": {
      "Python parent id": 5844, "Python id": 5848, "finished": true, "Ev Idx": 8748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907313.923, "dur": 0.176,
    "args": {
      "Python parent id": 5813, "Python id": 5849, "finished": true, "Ev Idx": 8749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907314.627, "dur": 0.913,
    "args": {
      "Python parent id": 5813, "Python id": 5850, "finished": true, "Ev Idx": 8750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907315.031, "dur": 0.464,
    "args": {
      "Python parent id": 5850, "Python id": 5851, "finished": true, "Ev Idx": 8751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907315.892, "dur": 2.603,
    "args": {
      "Python parent id": 5812, "Python id": 5852, "finished": true, "Ev Idx": 8752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907319.420, "dur": 84.130,
    "args": {
      "Python parent id": 12, "Python id": 5853, "finished": true, "Ev Idx": 8753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907319.701, "dur": 0.418,
    "args": {
      "Python parent id": 5853, "Python id": 5854, "finished": true, "Ev Idx": 8754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907319.967, "dur": 0.110,
    "args": {
      "Python parent id": 5854, "Python id": 5855, "finished": true, "Ev Idx": 8755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907320.407, "dur": 0.381,
    "args": {
      "Python parent id": 5853, "Python id": 5856, "finished": true, "Ev Idx": 8756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907320.980, "dur": 0.263,
    "args": {
      "Python parent id": 5853, "Python id": 5857, "finished": true, "Ev Idx": 8757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907321.474, "dur": 1.464,
    "args": {
      "Python parent id": 5853, "Python id": 5858, "finished": true, "Ev Idx": 8758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907323.437, "dur": 79.988,
    "args": {
      "Python parent id": 5853, "Python id": 5859, "finished": true, "Ev Idx": 8759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907324.181, "dur": 79.049,
    "args": {
      "Python parent id": 5859, "Python id": 5860, "finished": true, "Ev Idx": 8760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907324.459, "dur": 78.714,
    "args": {
      "Python parent id": 5860, "Python id": 5861, "finished": true, "Ev Idx": 8761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907329.177, "dur": 72.701,
    "args": {
      "Python parent id": 5861, "Python id": 5862, "finished": true, "Ev Idx": 8762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907329.497, "dur": 72.249,
    "args": {
      "Python parent id": 5862, "Python id": 5863, "finished": true, "Ev Idx": 8763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907330.015, "dur": 71.610,
    "args": {
      "Python parent id": 5863, "Python id": 5864, "finished": true, "Ev Idx": 8764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907330.541, "dur": 0.376,
    "args": {
      "Python parent id": 5864, "Python id": 5865, "finished": true, "Ev Idx": 8765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907330.788, "dur": 0.074,
    "args": {
      "Python parent id": 5865, "Python id": 5866, "finished": true, "Ev Idx": 8766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907331.167, "dur": 0.325,
    "args": {
      "Python parent id": 5864, "Python id": 5867, "finished": true, "Ev Idx": 8767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907331.382, "dur": 0.061,
    "args": {
      "Python parent id": 5867, "Python id": 5868, "finished": true, "Ev Idx": 8768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907331.786, "dur": 69.761,
    "args": {
      "Python parent id": 5864, "Python id": 5869, "finished": true, "Ev Idx": 8769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907332.431, "dur": 67.958,
    "args": {
      "Python parent id": 5869, "Python id": 5870, "finished": true, "Ev Idx": 8770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907339.026, "dur": 15.802,
    "args": {
      "Python parent id": 5870, "Python id": 5871, "finished": true, "Ev Idx": 8771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907339.147, "dur": 0.114,
    "args": {
      "Python parent id": 5871, "Python id": 5872, "finished": true, "Ev Idx": 8772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907339.521, "dur": 15.176,
    "args": {
      "Python parent id": 5871, "Python id": 5873, "finished": true, "Ev Idx": 8773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907339.645, "dur": 6.698,
    "args": {
      "Python parent id": 5873, "Python id": 5874, "finished": true, "Ev Idx": 8774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907363.301, "dur": 6.466,
    "args": {
      "Python parent id": 5870, "Python id": 5875, "finished": true, "Ev Idx": 8775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907378.322, "dur": 8.437,
    "args": {
      "Python parent id": 5870, "Python id": 5876, "finished": true, "Ev Idx": 8776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907404.633, "dur": 15.139,
    "args": {
      "Python parent id": 12, "Python id": 5877, "finished": true, "Ev Idx": 8777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907405.002, "dur": 11.368,
    "args": {
      "Python parent id": 5877, "Python id": 5878, "finished": true, "Ev Idx": 8778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907405.217, "dur": 0.498,
    "args": {
      "Python parent id": 5878, "Python id": 5879, "finished": true, "Ev Idx": 8779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907405.425, "dur": 0.227,
    "args": {
      "Python parent id": 5879, "Python id": 5880, "finished": true, "Ev Idx": 8780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907405.549, "dur": 0.064,
    "args": {
      "Python parent id": 5880, "Python id": 5881, "finished": true, "Ev Idx": 8781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907405.976, "dur": 8.363,
    "args": {
      "Python parent id": 5878, "Python id": 5882, "finished": true, "Ev Idx": 8782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907406.075, "dur": 0.059,
    "args": {
      "Python parent id": 5882, "Python id": 5883, "finished": true, "Ev Idx": 8783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907406.238, "dur": 0.059,
    "args": {
      "Python parent id": 5882, "Python id": 5884, "finished": true, "Ev Idx": 8784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907406.428, "dur": 0.059,
    "args": {
      "Python parent id": 5882, "Python id": 5885, "finished": true, "Ev Idx": 8785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907406.959, "dur": 0.038,
    "args": {
      "Python parent id": 5882, "Python id": 5886, "finished": true, "Ev Idx": 8786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907407.270, "dur": 0.079,
    "args": {
      "Python parent id": 5882, "Python id": 5887, "finished": true, "Ev Idx": 8787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907407.719, "dur": 6.552,
    "args": {
      "Python parent id": 5882, "Python id": 5888, "finished": true, "Ev Idx": 8788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907407.805, "dur": 0.065,
    "args": {
      "Python parent id": 5888, "Python id": 5889, "finished": true, "Ev Idx": 8789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907408.009, "dur": 0.057,
    "args": {
      "Python parent id": 5888, "Python id": 5890, "finished": true, "Ev Idx": 8790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907408.246, "dur": 0.061,
    "args": {
      "Python parent id": 5888, "Python id": 5891, "finished": true, "Ev Idx": 8791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907408.645, "dur": 0.028,
    "args": {
      "Python parent id": 5888, "Python id": 5892, "finished": true, "Ev Idx": 8792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907408.946, "dur": 5.274,
    "args": {
      "Python parent id": 5888, "Python id": 5893, "finished": true, "Ev Idx": 8793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.172, "dur": 4.997,
    "args": {
      "Python parent id": 5893, "Python id": 5894, "finished": true, "Ev Idx": 8794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.289, "dur": 3.206,
    "args": {
      "Python parent id": 5894, "Python id": 5895, "finished": true, "Ev Idx": 8795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.458, "dur": 2.998,
    "args": {
      "Python parent id": 5895, "Python id": 5896, "finished": true, "Ev Idx": 8796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.573, "dur": 0.265,
    "args": {
      "Python parent id": 5896, "Python id": 5897, "finished": true, "Ev Idx": 8797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.715, "dur": 0.063,
    "args": {
      "Python parent id": 5897, "Python id": 5898, "finished": true, "Ev Idx": 8798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907409.992, "dur": 2.103,
    "args": {
      "Python parent id": 5896, "Python id": 5899, "finished": true, "Ev Idx": 8799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907410.197, "dur": 1.828,
    "args": {
      "Python parent id": 5899, "Python id": 5900, "finished": true, "Ev Idx": 8800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907410.407, "dur": 1.570,
    "args": {
      "Python parent id": 5900, "Python id": 5901, "finished": true, "Ev Idx": 8801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907410.602, "dur": 1.133,
    "args": {
      "Python parent id": 5901, "Python id": 5902, "finished": true, "Ev Idx": 8802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907410.784, "dur": 0.475,
    "args": {
      "Python parent id": 5902, "Python id": 5903, "finished": true, "Ev Idx": 8803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907410.866, "dur": 0.050,
    "args": {
      "Python parent id": 5903, "Python id": 5904, "finished": true, "Ev Idx": 8804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907411.073, "dur": 0.133,
    "args": {
      "Python parent id": 5903, "Python id": 5905, "finished": true, "Ev Idx": 8805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907412.278, "dur": 0.136,
    "args": {
      "Python parent id": 5896, "Python id": 5906, "finished": true, "Ev Idx": 8806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907412.645, "dur": 0.100,
    "args": {
      "Python parent id": 5894, "Python id": 5907, "finished": true, "Ev Idx": 8807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.028, "dur": 1.075,
    "args": {
      "Python parent id": 5894, "Python id": 5908, "finished": true, "Ev Idx": 8808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.182, "dur": 0.881,
    "args": {
      "Python parent id": 5908, "Python id": 5909, "finished": true, "Ev Idx": 8809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.318, "dur": 0.359,
    "args": {
      "Python parent id": 5909, "Python id": 5910, "finished": true, "Ev Idx": 8810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.446, "dur": 0.178,
    "args": {
      "Python parent id": 5910, "Python id": 5911, "finished": true, "Ev Idx": 8811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.539, "dur": 0.056,
    "args": {
      "Python parent id": 5911, "Python id": 5912, "finished": true, "Ev Idx": 8812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907413.794, "dur": 0.239,
    "args": {
      "Python parent id": 5909, "Python id": 5913, "finished": true, "Ev Idx": 8813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907414.457, "dur": 0.192,
    "args": {
      "Python parent id": 5878, "Python id": 5914, "finished": true, "Ev Idx": 8814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907415.195, "dur": 1.024,
    "args": {
      "Python parent id": 5878, "Python id": 5915, "finished": true, "Ev Idx": 8815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907415.670, "dur": 0.501,
    "args": {
      "Python parent id": 5915, "Python id": 5916, "finished": true, "Ev Idx": 8816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907416.638, "dur": 3.050,
    "args": {
      "Python parent id": 5877, "Python id": 5917, "finished": true, "Ev Idx": 8817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907420.294, "dur": 6.996,
    "args": {
      "Python parent id": 12, "Python id": 5918, "finished": true, "Ev Idx": 8818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907427.776, "dur": 0.099,
    "args": {
      "Python parent id": 12, "Python id": 5919, "finished": true, "Ev Idx": 8819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907428.421, "dur": 0.186,
    "args": {
      "Python parent id": 12, "Python id": 5920, "finished": true, "Ev Idx": 8820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907429.094, "dur": 0.070,
    "args": {
      "Python parent id": 12, "Python id": 5921, "finished": true, "Ev Idx": 8821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907429.546, "dur": 8.625,
    "args": {
      "Python parent id": 12, "Python id": 5922, "finished": true, "Ev Idx": 8822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907439.307, "dur": 13.801,
    "args": {
      "Python parent id": 12, "Python id": 5923, "finished": true, "Ev Idx": 8823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907439.603, "dur": 10.418,
    "args": {
      "Python parent id": 5923, "Python id": 5924, "finished": true, "Ev Idx": 8824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907439.755, "dur": 0.441,
    "args": {
      "Python parent id": 5924, "Python id": 5925, "finished": true, "Ev Idx": 8825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907439.911, "dur": 0.213,
    "args": {
      "Python parent id": 5925, "Python id": 5926, "finished": true, "Ev Idx": 8826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907440.007, "dur": 0.085,
    "args": {
      "Python parent id": 5926, "Python id": 5927, "finished": true, "Ev Idx": 8827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907440.461, "dur": 7.896,
    "args": {
      "Python parent id": 5924, "Python id": 5928, "finished": true, "Ev Idx": 8828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907440.583, "dur": 0.064,
    "args": {
      "Python parent id": 5928, "Python id": 5929, "finished": true, "Ev Idx": 8829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907440.754, "dur": 0.054,
    "args": {
      "Python parent id": 5928, "Python id": 5930, "finished": true, "Ev Idx": 8830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907440.948, "dur": 0.062,
    "args": {
      "Python parent id": 5928, "Python id": 5931, "finished": true, "Ev Idx": 8831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907441.437, "dur": 0.051,
    "args": {
      "Python parent id": 5928, "Python id": 5932, "finished": true, "Ev Idx": 8832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907441.809, "dur": 0.074,
    "args": {
      "Python parent id": 5928, "Python id": 5933, "finished": true, "Ev Idx": 8833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907442.202, "dur": 6.089,
    "args": {
      "Python parent id": 5928, "Python id": 5934, "finished": true, "Ev Idx": 8834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907442.292, "dur": 0.060,
    "args": {
      "Python parent id": 5934, "Python id": 5935, "finished": true, "Ev Idx": 8835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907442.497, "dur": 0.056,
    "args": {
      "Python parent id": 5934, "Python id": 5936, "finished": true, "Ev Idx": 8836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907442.701, "dur": 0.078,
    "args": {
      "Python parent id": 5934, "Python id": 5937, "finished": true, "Ev Idx": 8837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907443.099, "dur": 0.037,
    "args": {
      "Python parent id": 5934, "Python id": 5938, "finished": true, "Ev Idx": 8838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907443.441, "dur": 4.774,
    "args": {
      "Python parent id": 5934, "Python id": 5939, "finished": true, "Ev Idx": 8839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907443.621, "dur": 4.540,
    "args": {
      "Python parent id": 5939, "Python id": 5940, "finished": true, "Ev Idx": 8840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907443.750, "dur": 2.827,
    "args": {
      "Python parent id": 5940, "Python id": 5941, "finished": true, "Ev Idx": 8841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907443.952, "dur": 2.580,
    "args": {
      "Python parent id": 5941, "Python id": 5942, "finished": true, "Ev Idx": 8842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.106, "dur": 0.225,
    "args": {
      "Python parent id": 5942, "Python id": 5943, "finished": true, "Ev Idx": 8843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.221, "dur": 0.060,
    "args": {
      "Python parent id": 5943, "Python id": 5944, "finished": true, "Ev Idx": 8844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.500, "dur": 1.755,
    "args": {
      "Python parent id": 5942, "Python id": 5945, "finished": true, "Ev Idx": 8845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.655, "dur": 1.542,
    "args": {
      "Python parent id": 5945, "Python id": 5946, "finished": true, "Ev Idx": 8846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.830, "dur": 1.308,
    "args": {
      "Python parent id": 5946, "Python id": 5947, "finished": true, "Ev Idx": 8847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907444.993, "dur": 0.928,
    "args": {
      "Python parent id": 5947, "Python id": 5948, "finished": true, "Ev Idx": 8848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907445.135, "dur": 0.424,
    "args": {
      "Python parent id": 5948, "Python id": 5949, "finished": true, "Ev Idx": 8849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907445.216, "dur": 0.042,
    "args": {
      "Python parent id": 5949, "Python id": 5950, "finished": true, "Ev Idx": 8850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907445.408, "dur": 0.100,
    "args": {
      "Python parent id": 5949, "Python id": 5951, "finished": true, "Ev Idx": 8851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907446.429, "dur": 0.068,
    "args": {
      "Python parent id": 5942, "Python id": 5952, "finished": true, "Ev Idx": 8852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907446.732, "dur": 0.082,
    "args": {
      "Python parent id": 5940, "Python id": 5953, "finished": true, "Ev Idx": 8853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.072, "dur": 1.045,
    "args": {
      "Python parent id": 5940, "Python id": 5954, "finished": true, "Ev Idx": 8854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.207, "dur": 0.868,
    "args": {
      "Python parent id": 5954, "Python id": 5955, "finished": true, "Ev Idx": 8855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.337, "dur": 0.368,
    "args": {
      "Python parent id": 5955, "Python id": 5956, "finished": true, "Ev Idx": 8856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.475, "dur": 0.172,
    "args": {
      "Python parent id": 5956, "Python id": 5957, "finished": true, "Ev Idx": 8857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.559, "dur": 0.057,
    "args": {
      "Python parent id": 5957, "Python id": 5958, "finished": true, "Ev Idx": 8858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907447.820, "dur": 0.226,
    "args": {
      "Python parent id": 5955, "Python id": 5959, "finished": true, "Ev Idx": 8859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907448.476, "dur": 0.143,
    "args": {
      "Python parent id": 5924, "Python id": 5960, "finished": true, "Ev Idx": 8860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907449.098, "dur": 0.810,
    "args": {
      "Python parent id": 5924, "Python id": 5961, "finished": true, "Ev Idx": 8861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907449.452, "dur": 0.408,
    "args": {
      "Python parent id": 5961, "Python id": 5962, "finished": true, "Ev Idx": 8862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907450.261, "dur": 2.794,
    "args": {
      "Python parent id": 5923, "Python id": 5963, "finished": true, "Ev Idx": 8863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907453.983, "dur": 82.449,
    "args": {
      "Python parent id": 12, "Python id": 5964, "finished": true, "Ev Idx": 8864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907454.318, "dur": 0.448,
    "args": {
      "Python parent id": 5964, "Python id": 5965, "finished": true, "Ev Idx": 8865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907454.592, "dur": 0.080,
    "args": {
      "Python parent id": 5965, "Python id": 5966, "finished": true, "Ev Idx": 8866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907455.060, "dur": 0.352,
    "args": {
      "Python parent id": 5964, "Python id": 5967, "finished": true, "Ev Idx": 8867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907455.618, "dur": 0.256,
    "args": {
      "Python parent id": 5964, "Python id": 5968, "finished": true, "Ev Idx": 8868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907456.072, "dur": 1.467,
    "args": {
      "Python parent id": 5964, "Python id": 5969, "finished": true, "Ev Idx": 8869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907458.123, "dur": 78.184,
    "args": {
      "Python parent id": 5964, "Python id": 5970, "finished": true, "Ev Idx": 8870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907458.827, "dur": 77.303,
    "args": {
      "Python parent id": 5970, "Python id": 5971, "finished": true, "Ev Idx": 8871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907459.105, "dur": 76.973,
    "args": {
      "Python parent id": 5971, "Python id": 5972, "finished": true, "Ev Idx": 8872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907464.017, "dur": 70.869,
    "args": {
      "Python parent id": 5972, "Python id": 5973, "finished": true, "Ev Idx": 8873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907464.275, "dur": 70.513,
    "args": {
      "Python parent id": 5973, "Python id": 5974, "finished": true, "Ev Idx": 8874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907464.747, "dur": 69.915,
    "args": {
      "Python parent id": 5974, "Python id": 5975, "finished": true, "Ev Idx": 8875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907465.287, "dur": 0.379,
    "args": {
      "Python parent id": 5975, "Python id": 5976, "finished": true, "Ev Idx": 8876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907465.548, "dur": 0.059,
    "args": {
      "Python parent id": 5976, "Python id": 5977, "finished": true, "Ev Idx": 8877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907465.929, "dur": 0.323,
    "args": {
      "Python parent id": 5975, "Python id": 5978, "finished": true, "Ev Idx": 8878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907466.165, "dur": 0.040,
    "args": {
      "Python parent id": 5978, "Python id": 5979, "finished": true, "Ev Idx": 8879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907466.587, "dur": 67.948,
    "args": {
      "Python parent id": 5975, "Python id": 5980, "finished": true, "Ev Idx": 8880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907467.166, "dur": 66.195,
    "args": {
      "Python parent id": 5980, "Python id": 5981, "finished": true, "Ev Idx": 8881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907473.857, "dur": 15.728,
    "args": {
      "Python parent id": 5981, "Python id": 5982, "finished": true, "Ev Idx": 8882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907474.001, "dur": 0.137,
    "args": {
      "Python parent id": 5982, "Python id": 5983, "finished": true, "Ev Idx": 8883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907474.402, "dur": 15.090,
    "args": {
      "Python parent id": 5982, "Python id": 5984, "finished": true, "Ev Idx": 8884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907474.521, "dur": 6.828,
    "args": {
      "Python parent id": 5984, "Python id": 5985, "finished": true, "Ev Idx": 8885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907497.600, "dur": 6.589,
    "args": {
      "Python parent id": 5981, "Python id": 5986, "finished": true, "Ev Idx": 8886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907511.948, "dur": 8.113,
    "args": {
      "Python parent id": 5981, "Python id": 5987, "finished": true, "Ev Idx": 8887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907537.519, "dur": 15.414,
    "args": {
      "Python parent id": 12, "Python id": 5988, "finished": true, "Ev Idx": 8888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907537.914, "dur": 11.519,
    "args": {
      "Python parent id": 5988, "Python id": 5989, "finished": true, "Ev Idx": 8889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907538.104, "dur": 0.518,
    "args": {
      "Python parent id": 5989, "Python id": 5990, "finished": true, "Ev Idx": 8890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907538.307, "dur": 0.228,
    "args": {
      "Python parent id": 5990, "Python id": 5991, "finished": true, "Ev Idx": 8891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907538.430, "dur": 0.068,
    "args": {
      "Python parent id": 5991, "Python id": 5992, "finished": true, "Ev Idx": 8892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907538.905, "dur": 8.489,
    "args": {
      "Python parent id": 5989, "Python id": 5993, "finished": true, "Ev Idx": 8893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907539.007, "dur": 0.059,
    "args": {
      "Python parent id": 5993, "Python id": 5994, "finished": true, "Ev Idx": 8894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907539.174, "dur": 0.058,
    "args": {
      "Python parent id": 5993, "Python id": 5995, "finished": true, "Ev Idx": 8895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907539.378, "dur": 0.058,
    "args": {
      "Python parent id": 5993, "Python id": 5996, "finished": true, "Ev Idx": 8896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907539.849, "dur": 0.045,
    "args": {
      "Python parent id": 5993, "Python id": 5997, "finished": true, "Ev Idx": 8897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907540.177, "dur": 0.061,
    "args": {
      "Python parent id": 5993, "Python id": 5998, "finished": true, "Ev Idx": 8898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907540.603, "dur": 6.736,
    "args": {
      "Python parent id": 5993, "Python id": 5999, "finished": true, "Ev Idx": 8899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907540.697, "dur": 0.058,
    "args": {
      "Python parent id": 5999, "Python id": 6000, "finished": true, "Ev Idx": 8900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907540.928, "dur": 0.052,
    "args": {
      "Python parent id": 5999, "Python id": 6001, "finished": true, "Ev Idx": 8901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907541.120, "dur": 0.079,
    "args": {
      "Python parent id": 5999, "Python id": 6002, "finished": true, "Ev Idx": 8902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907541.575, "dur": 0.048,
    "args": {
      "Python parent id": 5999, "Python id": 6003, "finished": true, "Ev Idx": 8903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907541.952, "dur": 5.294,
    "args": {
      "Python parent id": 5999, "Python id": 6004, "finished": true, "Ev Idx": 8904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907542.180, "dur": 4.994,
    "args": {
      "Python parent id": 6004, "Python id": 6005, "finished": true, "Ev Idx": 8905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907542.305, "dur": 3.150,
    "args": {
      "Python parent id": 6005, "Python id": 6006, "finished": true, "Ev Idx": 8906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907542.480, "dur": 2.909,
    "args": {
      "Python parent id": 6006, "Python id": 6007, "finished": true, "Ev Idx": 8907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907542.601, "dur": 0.275,
    "args": {
      "Python parent id": 6007, "Python id": 6008, "finished": true, "Ev Idx": 8908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907542.742, "dur": 0.059,
    "args": {
      "Python parent id": 6008, "Python id": 6009, "finished": true, "Ev Idx": 8909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.047, "dur": 1.997,
    "args": {
      "Python parent id": 6007, "Python id": 6010, "finished": true, "Ev Idx": 8910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.245, "dur": 1.726,
    "args": {
      "Python parent id": 6010, "Python id": 6011, "finished": true, "Ev Idx": 8911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.462, "dur": 1.445,
    "args": {
      "Python parent id": 6011, "Python id": 6012, "finished": true, "Ev Idx": 8912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.595, "dur": 1.104,
    "args": {
      "Python parent id": 6012, "Python id": 6013, "finished": true, "Ev Idx": 8913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.761, "dur": 0.488,
    "args": {
      "Python parent id": 6013, "Python id": 6014, "finished": true, "Ev Idx": 8914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907543.855, "dur": 0.041,
    "args": {
      "Python parent id": 6014, "Python id": 6015, "finished": true, "Ev Idx": 8915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907544.053, "dur": 0.146,
    "args": {
      "Python parent id": 6014, "Python id": 6016, "finished": true, "Ev Idx": 8916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907545.229, "dur": 0.125,
    "args": {
      "Python parent id": 6007, "Python id": 6017, "finished": true, "Ev Idx": 8917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907545.630, "dur": 0.107,
    "args": {
      "Python parent id": 6005, "Python id": 6018, "finished": true, "Ev Idx": 8918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907545.999, "dur": 1.135,
    "args": {
      "Python parent id": 6005, "Python id": 6019, "finished": true, "Ev Idx": 8919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907546.132, "dur": 0.958,
    "args": {
      "Python parent id": 6019, "Python id": 6020, "finished": true, "Ev Idx": 8920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907546.293, "dur": 0.360,
    "args": {
      "Python parent id": 6020, "Python id": 6021, "finished": true, "Ev Idx": 8921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907546.418, "dur": 0.180,
    "args": {
      "Python parent id": 6021, "Python id": 6022, "finished": true, "Ev Idx": 8922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907546.507, "dur": 0.062,
    "args": {
      "Python parent id": 6022, "Python id": 6023, "finished": true, "Ev Idx": 8923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907546.767, "dur": 0.295,
    "args": {
      "Python parent id": 6020, "Python id": 6024, "finished": true, "Ev Idx": 8924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907547.549, "dur": 0.258,
    "args": {
      "Python parent id": 5989, "Python id": 6025, "finished": true, "Ev Idx": 8925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907548.383, "dur": 0.924,
    "args": {
      "Python parent id": 5989, "Python id": 6026, "finished": true, "Ev Idx": 8926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907548.806, "dur": 0.451,
    "args": {
      "Python parent id": 6026, "Python id": 6027, "finished": true, "Ev Idx": 8927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907549.770, "dur": 3.095,
    "args": {
      "Python parent id": 5988, "Python id": 6028, "finished": true, "Ev Idx": 8928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907553.465, "dur": 6.970,
    "args": {
      "Python parent id": 12, "Python id": 6029, "finished": true, "Ev Idx": 8929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907560.871, "dur": 0.127,
    "args": {
      "Python parent id": 12, "Python id": 6030, "finished": true, "Ev Idx": 8930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907561.669, "dur": 0.169,
    "args": {
      "Python parent id": 12, "Python id": 6031, "finished": true, "Ev Idx": 8931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907562.307, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 6032, "finished": true, "Ev Idx": 8932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907562.736, "dur": 8.829,
    "args": {
      "Python parent id": 12, "Python id": 6033, "finished": true, "Ev Idx": 8933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907572.802, "dur": 13.997,
    "args": {
      "Python parent id": 12, "Python id": 6034, "finished": true, "Ev Idx": 8934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907573.087, "dur": 10.608,
    "args": {
      "Python parent id": 6034, "Python id": 6035, "finished": true, "Ev Idx": 8935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907573.228, "dur": 0.480,
    "args": {
      "Python parent id": 6035, "Python id": 6036, "finished": true, "Ev Idx": 8936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907573.400, "dur": 0.213,
    "args": {
      "Python parent id": 6036, "Python id": 6037, "finished": true, "Ev Idx": 8937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907573.495, "dur": 0.085,
    "args": {
      "Python parent id": 6037, "Python id": 6038, "finished": true, "Ev Idx": 8938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907574.006, "dur": 7.962,
    "args": {
      "Python parent id": 6035, "Python id": 6039, "finished": true, "Ev Idx": 8939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907574.087, "dur": 0.067,
    "args": {
      "Python parent id": 6039, "Python id": 6040, "finished": true, "Ev Idx": 8940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907574.271, "dur": 0.057,
    "args": {
      "Python parent id": 6039, "Python id": 6041, "finished": true, "Ev Idx": 8941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907574.472, "dur": 0.057,
    "args": {
      "Python parent id": 6039, "Python id": 6042, "finished": true, "Ev Idx": 8942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907574.898, "dur": 0.063,
    "args": {
      "Python parent id": 6039, "Python id": 6043, "finished": true, "Ev Idx": 8943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907575.242, "dur": 0.071,
    "args": {
      "Python parent id": 6039, "Python id": 6044, "finished": true, "Ev Idx": 8944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907575.636, "dur": 6.248,
    "args": {
      "Python parent id": 6039, "Python id": 6045, "finished": true, "Ev Idx": 8945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907575.744, "dur": 0.059,
    "args": {
      "Python parent id": 6045, "Python id": 6046, "finished": true, "Ev Idx": 8946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907575.985, "dur": 0.059,
    "args": {
      "Python parent id": 6045, "Python id": 6047, "finished": true, "Ev Idx": 8947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907576.243, "dur": 0.067,
    "args": {
      "Python parent id": 6045, "Python id": 6048, "finished": true, "Ev Idx": 8948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907576.637, "dur": 0.064,
    "args": {
      "Python parent id": 6045, "Python id": 6049, "finished": true, "Ev Idx": 8949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.001, "dur": 4.832,
    "args": {
      "Python parent id": 6045, "Python id": 6050, "finished": true, "Ev Idx": 8950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.219, "dur": 4.561,
    "args": {
      "Python parent id": 6050, "Python id": 6051, "finished": true, "Ev Idx": 8951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.387, "dur": 2.765,
    "args": {
      "Python parent id": 6051, "Python id": 6052, "finished": true, "Ev Idx": 8952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.576, "dur": 2.534,
    "args": {
      "Python parent id": 6052, "Python id": 6053, "finished": true, "Ev Idx": 8953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.718, "dur": 0.226,
    "args": {
      "Python parent id": 6053, "Python id": 6054, "finished": true, "Ev Idx": 8954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907577.833, "dur": 0.065,
    "args": {
      "Python parent id": 6054, "Python id": 6055, "finished": true, "Ev Idx": 8955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.114, "dur": 1.701,
    "args": {
      "Python parent id": 6053, "Python id": 6056, "finished": true, "Ev Idx": 8956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.264, "dur": 1.495,
    "args": {
      "Python parent id": 6056, "Python id": 6057, "finished": true, "Ev Idx": 8957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.420, "dur": 1.295,
    "args": {
      "Python parent id": 6057, "Python id": 6058, "finished": true, "Ev Idx": 8958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.573, "dur": 0.986,
    "args": {
      "Python parent id": 6058, "Python id": 6059, "finished": true, "Ev Idx": 8959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.723, "dur": 0.436,
    "args": {
      "Python parent id": 6059, "Python id": 6060, "finished": true, "Ev Idx": 8960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907578.799, "dur": 0.043,
    "args": {
      "Python parent id": 6060, "Python id": 6061, "finished": true, "Ev Idx": 8961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907579.019, "dur": 0.102,
    "args": {
      "Python parent id": 6060, "Python id": 6062, "finished": true, "Ev Idx": 8962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907579.987, "dur": 0.087,
    "args": {
      "Python parent id": 6053, "Python id": 6063, "finished": true, "Ev Idx": 8963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907580.329, "dur": 0.083,
    "args": {
      "Python parent id": 6051, "Python id": 6064, "finished": true, "Ev Idx": 8964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907580.678, "dur": 1.060,
    "args": {
      "Python parent id": 6051, "Python id": 6065, "finished": true, "Ev Idx": 8965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907580.803, "dur": 0.894,
    "args": {
      "Python parent id": 6065, "Python id": 6066, "finished": true, "Ev Idx": 8966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907580.937, "dur": 0.355,
    "args": {
      "Python parent id": 6066, "Python id": 6067, "finished": true, "Ev Idx": 8967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907581.071, "dur": 0.167,
    "args": {
      "Python parent id": 6067, "Python id": 6068, "finished": true, "Ev Idx": 8968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907581.156, "dur": 0.054,
    "args": {
      "Python parent id": 6068, "Python id": 6069, "finished": true, "Ev Idx": 8969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907581.408, "dur": 0.260,
    "args": {
      "Python parent id": 6066, "Python id": 6070, "finished": true, "Ev Idx": 8970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907582.079, "dur": 0.178,
    "args": {
      "Python parent id": 6035, "Python id": 6071, "finished": true, "Ev Idx": 8971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907582.760, "dur": 0.809,
    "args": {
      "Python parent id": 6035, "Python id": 6072, "finished": true, "Ev Idx": 8972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907583.131, "dur": 0.392,
    "args": {
      "Python parent id": 6072, "Python id": 6073, "finished": true, "Ev Idx": 8973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907583.952, "dur": 2.792,
    "args": {
      "Python parent id": 6034, "Python id": 6074, "finished": true, "Ev Idx": 8974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907587.669, "dur": 89.086,
    "args": {
      "Python parent id": 12, "Python id": 6075, "finished": true, "Ev Idx": 8975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907587.969, "dur": 0.415,
    "args": {
      "Python parent id": 6075, "Python id": 6076, "finished": true, "Ev Idx": 8976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907588.210, "dur": 0.076,
    "args": {
      "Python parent id": 6076, "Python id": 6077, "finished": true, "Ev Idx": 8977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907588.669, "dur": 0.283,
    "args": {
      "Python parent id": 6075, "Python id": 6078, "finished": true, "Ev Idx": 8978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907589.136, "dur": 0.249,
    "args": {
      "Python parent id": 6075, "Python id": 6079, "finished": true, "Ev Idx": 8979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907589.584, "dur": 1.411,
    "args": {
      "Python parent id": 6075, "Python id": 6080, "finished": true, "Ev Idx": 8980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907591.500, "dur": 85.156,
    "args": {
      "Python parent id": 6075, "Python id": 6081, "finished": true, "Ev Idx": 8981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907592.302, "dur": 84.138,
    "args": {
      "Python parent id": 6081, "Python id": 6082, "finished": true, "Ev Idx": 8982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907592.557, "dur": 83.799,
    "args": {
      "Python parent id": 6082, "Python id": 6083, "finished": true, "Ev Idx": 8983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907597.127, "dur": 78.073,
    "args": {
      "Python parent id": 6083, "Python id": 6084, "finished": true, "Ev Idx": 8984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907597.368, "dur": 77.783,
    "args": {
      "Python parent id": 6084, "Python id": 6085, "finished": true, "Ev Idx": 8985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907597.889, "dur": 77.124,
    "args": {
      "Python parent id": 6085, "Python id": 6086, "finished": true, "Ev Idx": 8986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907598.343, "dur": 0.393,
    "args": {
      "Python parent id": 6086, "Python id": 6087, "finished": true, "Ev Idx": 8987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907598.624, "dur": 0.045,
    "args": {
      "Python parent id": 6087, "Python id": 6088, "finished": true, "Ev Idx": 8988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907599.029, "dur": 0.288,
    "args": {
      "Python parent id": 6086, "Python id": 6089, "finished": true, "Ev Idx": 8989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907599.229, "dur": 0.038,
    "args": {
      "Python parent id": 6089, "Python id": 6090, "finished": true, "Ev Idx": 8990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907599.634, "dur": 75.186,
    "args": {
      "Python parent id": 6086, "Python id": 6091, "finished": true, "Ev Idx": 8991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907600.271, "dur": 73.403,
    "args": {
      "Python parent id": 6091, "Python id": 6092, "finished": true, "Ev Idx": 8992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907606.732, "dur": 15.389,
    "args": {
      "Python parent id": 6092, "Python id": 6093, "finished": true, "Ev Idx": 8993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907606.860, "dur": 0.138,
    "args": {
      "Python parent id": 6093, "Python id": 6094, "finished": true, "Ev Idx": 8994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907607.243, "dur": 14.733,
    "args": {
      "Python parent id": 6093, "Python id": 6095, "finished": true, "Ev Idx": 8995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907607.357, "dur": 6.547,
    "args": {
      "Python parent id": 6095, "Python id": 6096, "finished": true, "Ev Idx": 8996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907633.091, "dur": 7.501,
    "args": {
      "Python parent id": 6092, "Python id": 6097, "finished": true, "Ev Idx": 8997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907649.884, "dur": 9.214,
    "args": {
      "Python parent id": 6092, "Python id": 6098, "finished": true, "Ev Idx": 8998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907677.919, "dur": 15.722,
    "args": {
      "Python parent id": 12, "Python id": 6099, "finished": true, "Ev Idx": 8999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907678.272, "dur": 11.855,
    "args": {
      "Python parent id": 6099, "Python id": 6100, "finished": true, "Ev Idx": 9000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907678.451, "dur": 0.454,
    "args": {
      "Python parent id": 6100, "Python id": 6101, "finished": true, "Ev Idx": 9001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907678.615, "dur": 0.229,
    "args": {
      "Python parent id": 6101, "Python id": 6102, "finished": true, "Ev Idx": 9002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907678.732, "dur": 0.069,
    "args": {
      "Python parent id": 6102, "Python id": 6103, "finished": true, "Ev Idx": 9003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907679.173, "dur": 8.771,
    "args": {
      "Python parent id": 6100, "Python id": 6104, "finished": true, "Ev Idx": 9004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907679.300, "dur": 0.065,
    "args": {
      "Python parent id": 6104, "Python id": 6105, "finished": true, "Ev Idx": 9005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907679.468, "dur": 0.059,
    "args": {
      "Python parent id": 6104, "Python id": 6106, "finished": true, "Ev Idx": 9006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907679.660, "dur": 0.053,
    "args": {
      "Python parent id": 6104, "Python id": 6107, "finished": true, "Ev Idx": 9007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907680.190, "dur": 0.026,
    "args": {
      "Python parent id": 6104, "Python id": 6108, "finished": true, "Ev Idx": 9008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907680.518, "dur": 0.090,
    "args": {
      "Python parent id": 6104, "Python id": 6109, "finished": true, "Ev Idx": 9009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907680.965, "dur": 6.925,
    "args": {
      "Python parent id": 6104, "Python id": 6110, "finished": true, "Ev Idx": 9010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907681.064, "dur": 0.059,
    "args": {
      "Python parent id": 6110, "Python id": 6111, "finished": true, "Ev Idx": 9011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907681.278, "dur": 0.054,
    "args": {
      "Python parent id": 6110, "Python id": 6112, "finished": true, "Ev Idx": 9012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907681.526, "dur": 0.063,
    "args": {
      "Python parent id": 6110, "Python id": 6113, "finished": true, "Ev Idx": 9013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907681.940, "dur": 0.033,
    "args": {
      "Python parent id": 6110, "Python id": 6114, "finished": true, "Ev Idx": 9014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907682.268, "dur": 5.540,
    "args": {
      "Python parent id": 6110, "Python id": 6115, "finished": true, "Ev Idx": 9015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907682.594, "dur": 5.140,
    "args": {
      "Python parent id": 6115, "Python id": 6116, "finished": true, "Ev Idx": 9016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907682.722, "dur": 3.314,
    "args": {
      "Python parent id": 6116, "Python id": 6117, "finished": true, "Ev Idx": 9017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907682.921, "dur": 3.035,
    "args": {
      "Python parent id": 6117, "Python id": 6118, "finished": true, "Ev Idx": 9018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907683.052, "dur": 0.287,
    "args": {
      "Python parent id": 6118, "Python id": 6119, "finished": true, "Ev Idx": 9019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907683.209, "dur": 0.062,
    "args": {
      "Python parent id": 6119, "Python id": 6120, "finished": true, "Ev Idx": 9020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907683.482, "dur": 2.075,
    "args": {
      "Python parent id": 6118, "Python id": 6121, "finished": true, "Ev Idx": 9021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907683.674, "dur": 1.788,
    "args": {
      "Python parent id": 6121, "Python id": 6122, "finished": true, "Ev Idx": 9022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907683.900, "dur": 1.499,
    "args": {
      "Python parent id": 6122, "Python id": 6123, "finished": true, "Ev Idx": 9023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907684.063, "dur": 1.114,
    "args": {
      "Python parent id": 6123, "Python id": 6124, "finished": true, "Ev Idx": 9024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907684.244, "dur": 0.477,
    "args": {
      "Python parent id": 6124, "Python id": 6125, "finished": true, "Ev Idx": 9025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907684.341, "dur": 0.050,
    "args": {
      "Python parent id": 6125, "Python id": 6126, "finished": true, "Ev Idx": 9026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907684.550, "dur": 0.133,
    "args": {
      "Python parent id": 6125, "Python id": 6127, "finished": true, "Ev Idx": 9027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907685.733, "dur": 0.180,
    "args": {
      "Python parent id": 6118, "Python id": 6128, "finished": true, "Ev Idx": 9028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907686.190, "dur": 0.073,
    "args": {
      "Python parent id": 6116, "Python id": 6129, "finished": true, "Ev Idx": 9029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907686.560, "dur": 1.135,
    "args": {
      "Python parent id": 6116, "Python id": 6130, "finished": true, "Ev Idx": 9030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907686.716, "dur": 0.942,
    "args": {
      "Python parent id": 6130, "Python id": 6131, "finished": true, "Ev Idx": 9031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907686.840, "dur": 0.375,
    "args": {
      "Python parent id": 6131, "Python id": 6132, "finished": true, "Ev Idx": 9032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907686.969, "dur": 0.173,
    "args": {
      "Python parent id": 6132, "Python id": 6133, "finished": true, "Ev Idx": 9033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907687.067, "dur": 0.047,
    "args": {
      "Python parent id": 6133, "Python id": 6134, "finished": true, "Ev Idx": 9034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907687.335, "dur": 0.276,
    "args": {
      "Python parent id": 6131, "Python id": 6135, "finished": true, "Ev Idx": 9035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907688.086, "dur": 0.258,
    "args": {
      "Python parent id": 6100, "Python id": 6136, "finished": true, "Ev Idx": 9036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907688.904, "dur": 1.059,
    "args": {
      "Python parent id": 6100, "Python id": 6137, "finished": true, "Ev Idx": 9037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907689.395, "dur": 0.519,
    "args": {
      "Python parent id": 6137, "Python id": 6138, "finished": true, "Ev Idx": 9038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907690.406, "dur": 3.144,
    "args": {
      "Python parent id": 6099, "Python id": 6139, "finished": true, "Ev Idx": 9039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907694.165, "dur": 7.425,
    "args": {
      "Python parent id": 12, "Python id": 6140, "finished": true, "Ev Idx": 9040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907702.018, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 6141, "finished": true, "Ev Idx": 9041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907702.658, "dur": 0.188,
    "args": {
      "Python parent id": 12, "Python id": 6142, "finished": true, "Ev Idx": 9042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907703.311, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 6143, "finished": true, "Ev Idx": 9043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907703.751, "dur": 9.625,
    "args": {
      "Python parent id": 12, "Python id": 6144, "finished": true, "Ev Idx": 9044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907714.598, "dur": 15.068,
    "args": {
      "Python parent id": 12, "Python id": 6145, "finished": true, "Ev Idx": 9045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907714.928, "dur": 11.746,
    "args": {
      "Python parent id": 6145, "Python id": 6146, "finished": true, "Ev Idx": 9046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907715.100, "dur": 0.449,
    "args": {
      "Python parent id": 6146, "Python id": 6147, "finished": true, "Ev Idx": 9047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907715.239, "dur": 0.237,
    "args": {
      "Python parent id": 6147, "Python id": 6148, "finished": true, "Ev Idx": 9048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907715.328, "dur": 0.099,
    "args": {
      "Python parent id": 6148, "Python id": 6149, "finished": true, "Ev Idx": 9049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907715.809, "dur": 9.020,
    "args": {
      "Python parent id": 6146, "Python id": 6150, "finished": true, "Ev Idx": 9050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907715.893, "dur": 0.060,
    "args": {
      "Python parent id": 6150, "Python id": 6151, "finished": true, "Ev Idx": 9051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907716.061, "dur": 0.057,
    "args": {
      "Python parent id": 6150, "Python id": 6152, "finished": true, "Ev Idx": 9052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907716.301, "dur": 0.056,
    "args": {
      "Python parent id": 6150, "Python id": 6153, "finished": true, "Ev Idx": 9053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907716.745, "dur": 0.034,
    "args": {
      "Python parent id": 6150, "Python id": 6154, "finished": true, "Ev Idx": 9054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907717.093, "dur": 0.068,
    "args": {
      "Python parent id": 6150, "Python id": 6155, "finished": true, "Ev Idx": 9055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907717.446, "dur": 7.314,
    "args": {
      "Python parent id": 6150, "Python id": 6156, "finished": true, "Ev Idx": 9056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907717.536, "dur": 0.062,
    "args": {
      "Python parent id": 6156, "Python id": 6157, "finished": true, "Ev Idx": 9057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907717.767, "dur": 0.060,
    "args": {
      "Python parent id": 6156, "Python id": 6158, "finished": true, "Ev Idx": 9058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907718.019, "dur": 0.064,
    "args": {
      "Python parent id": 6156, "Python id": 6159, "finished": true, "Ev Idx": 9059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907718.429, "dur": 0.035,
    "args": {
      "Python parent id": 6156, "Python id": 6160, "finished": true, "Ev Idx": 9060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907718.733, "dur": 5.970,
    "args": {
      "Python parent id": 6156, "Python id": 6161, "finished": true, "Ev Idx": 9061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907718.948, "dur": 5.700,
    "args": {
      "Python parent id": 6161, "Python id": 6162, "finished": true, "Ev Idx": 9062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907720.287, "dur": 2.739,
    "args": {
      "Python parent id": 6162, "Python id": 6163, "finished": true, "Ev Idx": 9063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907720.466, "dur": 2.500,
    "args": {
      "Python parent id": 6163, "Python id": 6164, "finished": true, "Ev Idx": 9064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907720.591, "dur": 0.235,
    "args": {
      "Python parent id": 6164, "Python id": 6165, "finished": true, "Ev Idx": 9065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907720.720, "dur": 0.054,
    "args": {
      "Python parent id": 6165, "Python id": 6166, "finished": true, "Ev Idx": 9066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907720.966, "dur": 1.732,
    "args": {
      "Python parent id": 6164, "Python id": 6167, "finished": true, "Ev Idx": 9067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.148, "dur": 1.493,
    "args": {
      "Python parent id": 6167, "Python id": 6168, "finished": true, "Ev Idx": 9068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.331, "dur": 1.268,
    "args": {
      "Python parent id": 6168, "Python id": 6169, "finished": true, "Ev Idx": 9069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.460, "dur": 0.977,
    "args": {
      "Python parent id": 6169, "Python id": 6170, "finished": true, "Ev Idx": 9070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.612, "dur": 0.462,
    "args": {
      "Python parent id": 6170, "Python id": 6171, "finished": true, "Ev Idx": 9071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.703, "dur": 0.041,
    "args": {
      "Python parent id": 6171, "Python id": 6172, "finished": true, "Ev Idx": 9072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907721.906, "dur": 0.127,
    "args": {
      "Python parent id": 6171, "Python id": 6173, "finished": true, "Ev Idx": 9073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907722.872, "dur": 0.060,
    "args": {
      "Python parent id": 6164, "Python id": 6174, "finished": true, "Ev Idx": 9074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907723.163, "dur": 0.094,
    "args": {
      "Python parent id": 6162, "Python id": 6175, "finished": true, "Ev Idx": 9075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907723.521, "dur": 1.086,
    "args": {
      "Python parent id": 6162, "Python id": 6176, "finished": true, "Ev Idx": 9076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907723.678, "dur": 0.890,
    "args": {
      "Python parent id": 6176, "Python id": 6177, "finished": true, "Ev Idx": 9077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907723.805, "dur": 0.355,
    "args": {
      "Python parent id": 6177, "Python id": 6178, "finished": true, "Ev Idx": 9078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907723.931, "dur": 0.172,
    "args": {
      "Python parent id": 6178, "Python id": 6179, "finished": true, "Ev Idx": 9079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907724.019, "dur": 0.053,
    "args": {
      "Python parent id": 6179, "Python id": 6180, "finished": true, "Ev Idx": 9080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907724.279, "dur": 0.261,
    "args": {
      "Python parent id": 6177, "Python id": 6181, "finished": true, "Ev Idx": 9081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907724.976, "dur": 0.173,
    "args": {
      "Python parent id": 6146, "Python id": 6182, "finished": true, "Ev Idx": 9082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907725.691, "dur": 0.830,
    "args": {
      "Python parent id": 6146, "Python id": 6183, "finished": true, "Ev Idx": 9083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907726.068, "dur": 0.405,
    "args": {
      "Python parent id": 6183, "Python id": 6184, "finished": true, "Ev Idx": 9084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907726.940, "dur": 2.654,
    "args": {
      "Python parent id": 6145, "Python id": 6185, "finished": true, "Ev Idx": 9085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907730.484, "dur": 87.548,
    "args": {
      "Python parent id": 12, "Python id": 6186, "finished": true, "Ev Idx": 9086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907730.743, "dur": 0.441,
    "args": {
      "Python parent id": 6186, "Python id": 6187, "finished": true, "Ev Idx": 9087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907731.011, "dur": 0.113,
    "args": {
      "Python parent id": 6187, "Python id": 6188, "finished": true, "Ev Idx": 9088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907731.462, "dur": 0.336,
    "args": {
      "Python parent id": 6186, "Python id": 6189, "finished": true, "Ev Idx": 9089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907731.999, "dur": 0.242,
    "args": {
      "Python parent id": 6186, "Python id": 6190, "finished": true, "Ev Idx": 9090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907732.449, "dur": 2.224,
    "args": {
      "Python parent id": 6186, "Python id": 6191, "finished": true, "Ev Idx": 9091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907735.201, "dur": 82.720,
    "args": {
      "Python parent id": 6186, "Python id": 6192, "finished": true, "Ev Idx": 9092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907736.082, "dur": 81.700,
    "args": {
      "Python parent id": 6192, "Python id": 6193, "finished": true, "Ev Idx": 9093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907736.344, "dur": 81.361,
    "args": {
      "Python parent id": 6193, "Python id": 6194, "finished": true, "Ev Idx": 9094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907741.251, "dur": 75.299,
    "args": {
      "Python parent id": 6194, "Python id": 6195, "finished": true, "Ev Idx": 9095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907741.549, "dur": 74.922,
    "args": {
      "Python parent id": 6195, "Python id": 6196, "finished": true, "Ev Idx": 9096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907742.078, "dur": 74.174,
    "args": {
      "Python parent id": 6196, "Python id": 6197, "finished": true, "Ev Idx": 9097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907742.548, "dur": 0.357,
    "args": {
      "Python parent id": 6197, "Python id": 6198, "finished": true, "Ev Idx": 9098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907742.810, "dur": 0.040,
    "args": {
      "Python parent id": 6198, "Python id": 6199, "finished": true, "Ev Idx": 9099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907743.199, "dur": 0.322,
    "args": {
      "Python parent id": 6197, "Python id": 6200, "finished": true, "Ev Idx": 9100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907743.414, "dur": 0.066,
    "args": {
      "Python parent id": 6200, "Python id": 6201, "finished": true, "Ev Idx": 9101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907743.843, "dur": 72.321,
    "args": {
      "Python parent id": 6197, "Python id": 6202, "finished": true, "Ev Idx": 9102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907744.462, "dur": 70.500,
    "args": {
      "Python parent id": 6202, "Python id": 6203, "finished": true, "Ev Idx": 9103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907751.438, "dur": 17.498,
    "args": {
      "Python parent id": 6203, "Python id": 6204, "finished": true, "Ev Idx": 9104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907751.559, "dur": 0.136,
    "args": {
      "Python parent id": 6204, "Python id": 6205, "finished": true, "Ev Idx": 9105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907751.942, "dur": 16.861,
    "args": {
      "Python parent id": 6204, "Python id": 6206, "finished": true, "Ev Idx": 9106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907752.097, "dur": 8.008,
    "args": {
      "Python parent id": 6206, "Python id": 6207, "finished": true, "Ev Idx": 9107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907777.657, "dur": 6.649,
    "args": {
      "Python parent id": 6203, "Python id": 6208, "finished": true, "Ev Idx": 9108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907792.311, "dur": 8.564,
    "args": {
      "Python parent id": 6203, "Python id": 6209, "finished": true, "Ev Idx": 9109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907819.205, "dur": 14.777,
    "args": {
      "Python parent id": 12, "Python id": 6210, "finished": true, "Ev Idx": 9110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907819.539, "dur": 11.206,
    "args": {
      "Python parent id": 6210, "Python id": 6211, "finished": true, "Ev Idx": 9111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907819.701, "dur": 0.424,
    "args": {
      "Python parent id": 6211, "Python id": 6212, "finished": true, "Ev Idx": 9112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907819.840, "dur": 0.204,
    "args": {
      "Python parent id": 6212, "Python id": 6213, "finished": true, "Ev Idx": 9113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907819.935, "dur": 0.077,
    "args": {
      "Python parent id": 6213, "Python id": 6214, "finished": true, "Ev Idx": 9114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907820.389, "dur": 8.388,
    "args": {
      "Python parent id": 6211, "Python id": 6215, "finished": true, "Ev Idx": 9115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907820.506, "dur": 0.108,
    "args": {
      "Python parent id": 6215, "Python id": 6216, "finished": true, "Ev Idx": 9116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907820.704, "dur": 0.055,
    "args": {
      "Python parent id": 6215, "Python id": 6217, "finished": true, "Ev Idx": 9117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907820.901, "dur": 0.049,
    "args": {
      "Python parent id": 6215, "Python id": 6218, "finished": true, "Ev Idx": 9118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907821.376, "dur": 0.056,
    "args": {
      "Python parent id": 6215, "Python id": 6219, "finished": true, "Ev Idx": 9119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907821.692, "dur": 0.064,
    "args": {
      "Python parent id": 6215, "Python id": 6220, "finished": true, "Ev Idx": 9120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907822.034, "dur": 6.694,
    "args": {
      "Python parent id": 6215, "Python id": 6221, "finished": true, "Ev Idx": 9121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907822.129, "dur": 0.061,
    "args": {
      "Python parent id": 6221, "Python id": 6222, "finished": true, "Ev Idx": 9122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907822.331, "dur": 0.047,
    "args": {
      "Python parent id": 6221, "Python id": 6223, "finished": true, "Ev Idx": 9123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907822.561, "dur": 0.065,
    "args": {
      "Python parent id": 6221, "Python id": 6224, "finished": true, "Ev Idx": 9124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907822.899, "dur": 0.032,
    "args": {
      "Python parent id": 6221, "Python id": 6225, "finished": true, "Ev Idx": 9125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907823.195, "dur": 5.437,
    "args": {
      "Python parent id": 6221, "Python id": 6226, "finished": true, "Ev Idx": 9126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907823.461, "dur": 5.098,
    "args": {
      "Python parent id": 6226, "Python id": 6227, "finished": true, "Ev Idx": 9127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907823.580, "dur": 3.261,
    "args": {
      "Python parent id": 6227, "Python id": 6228, "finished": true, "Ev Idx": 9128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907823.796, "dur": 2.976,
    "args": {
      "Python parent id": 6228, "Python id": 6229, "finished": true, "Ev Idx": 9129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907823.922, "dur": 0.285,
    "args": {
      "Python parent id": 6229, "Python id": 6230, "finished": true, "Ev Idx": 9130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907824.073, "dur": 0.097,
    "args": {
      "Python parent id": 6230, "Python id": 6231, "finished": true, "Ev Idx": 9131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907824.344, "dur": 2.072,
    "args": {
      "Python parent id": 6229, "Python id": 6232, "finished": true, "Ev Idx": 9132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907824.563, "dur": 1.782,
    "args": {
      "Python parent id": 6232, "Python id": 6233, "finished": true, "Ev Idx": 9133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907824.747, "dur": 1.556,
    "args": {
      "Python parent id": 6233, "Python id": 6234, "finished": true, "Ev Idx": 9134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907824.900, "dur": 1.191,
    "args": {
      "Python parent id": 6234, "Python id": 6235, "finished": true, "Ev Idx": 9135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907825.072, "dur": 0.553,
    "args": {
      "Python parent id": 6235, "Python id": 6236, "finished": true, "Ev Idx": 9136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907825.154, "dur": 0.063,
    "args": {
      "Python parent id": 6236, "Python id": 6237, "finished": true, "Ev Idx": 9137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907825.394, "dur": 0.168,
    "args": {
      "Python parent id": 6236, "Python id": 6238, "finished": true, "Ev Idx": 9138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907826.617, "dur": 0.108,
    "args": {
      "Python parent id": 6229, "Python id": 6239, "finished": true, "Ev Idx": 9139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907826.993, "dur": 0.097,
    "args": {
      "Python parent id": 6227, "Python id": 6240, "finished": true, "Ev Idx": 9140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907827.386, "dur": 1.131,
    "args": {
      "Python parent id": 6227, "Python id": 6241, "finished": true, "Ev Idx": 9141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907827.516, "dur": 0.939,
    "args": {
      "Python parent id": 6241, "Python id": 6242, "finished": true, "Ev Idx": 9142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907827.672, "dur": 0.364,
    "args": {
      "Python parent id": 6242, "Python id": 6243, "finished": true, "Ev Idx": 9143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907827.805, "dur": 0.176,
    "args": {
      "Python parent id": 6243, "Python id": 6244, "finished": true, "Ev Idx": 9144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907827.896, "dur": 0.056,
    "args": {
      "Python parent id": 6244, "Python id": 6245, "finished": true, "Ev Idx": 9145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907828.157, "dur": 0.249,
    "args": {
      "Python parent id": 6242, "Python id": 6246, "finished": true, "Ev Idx": 9146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907828.917, "dur": 0.240,
    "args": {
      "Python parent id": 6211, "Python id": 6247, "finished": true, "Ev Idx": 9147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907829.746, "dur": 0.879,
    "args": {
      "Python parent id": 6211, "Python id": 6248, "finished": true, "Ev Idx": 9148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907830.127, "dur": 0.418,
    "args": {
      "Python parent id": 6248, "Python id": 6249, "finished": true, "Ev Idx": 9149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907831.053, "dur": 2.852,
    "args": {
      "Python parent id": 6210, "Python id": 6250, "finished": true, "Ev Idx": 9150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907834.550, "dur": 7.374,
    "args": {
      "Python parent id": 12, "Python id": 6251, "finished": true, "Ev Idx": 9151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907842.395, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 6252, "finished": true, "Ev Idx": 9152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907843.151, "dur": 0.232,
    "args": {
      "Python parent id": 12, "Python id": 6253, "finished": true, "Ev Idx": 9153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907843.874, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 6254, "finished": true, "Ev Idx": 9154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907844.341, "dur": 9.023,
    "args": {
      "Python parent id": 12, "Python id": 6255, "finished": true, "Ev Idx": 9155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907854.432, "dur": 13.902,
    "args": {
      "Python parent id": 12, "Python id": 6256, "finished": true, "Ev Idx": 9156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907854.767, "dur": 10.256,
    "args": {
      "Python parent id": 6256, "Python id": 6257, "finished": true, "Ev Idx": 9157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907854.904, "dur": 0.442,
    "args": {
      "Python parent id": 6257, "Python id": 6258, "finished": true, "Ev Idx": 9158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907855.051, "dur": 0.223,
    "args": {
      "Python parent id": 6258, "Python id": 6259, "finished": true, "Ev Idx": 9159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907855.146, "dur": 0.096,
    "args": {
      "Python parent id": 6259, "Python id": 6260, "finished": true, "Ev Idx": 9160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907855.601, "dur": 7.673,
    "args": {
      "Python parent id": 6257, "Python id": 6261, "finished": true, "Ev Idx": 9161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907855.679, "dur": 0.065,
    "args": {
      "Python parent id": 6261, "Python id": 6262, "finished": true, "Ev Idx": 9162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907855.835, "dur": 0.062,
    "args": {
      "Python parent id": 6261, "Python id": 6263, "finished": true, "Ev Idx": 9163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907856.020, "dur": 0.055,
    "args": {
      "Python parent id": 6261, "Python id": 6264, "finished": true, "Ev Idx": 9164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907856.485, "dur": 0.067,
    "args": {
      "Python parent id": 6261, "Python id": 6265, "finished": true, "Ev Idx": 9165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907856.822, "dur": 0.070,
    "args": {
      "Python parent id": 6261, "Python id": 6266, "finished": true, "Ev Idx": 9166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907857.204, "dur": 6.013,
    "args": {
      "Python parent id": 6261, "Python id": 6267, "finished": true, "Ev Idx": 9167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907857.297, "dur": 0.059,
    "args": {
      "Python parent id": 6267, "Python id": 6268, "finished": true, "Ev Idx": 9168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907857.525, "dur": 0.058,
    "args": {
      "Python parent id": 6267, "Python id": 6269, "finished": true, "Ev Idx": 9169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907857.757, "dur": 0.066,
    "args": {
      "Python parent id": 6267, "Python id": 6270, "finished": true, "Ev Idx": 9170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907858.160, "dur": 0.028,
    "args": {
      "Python parent id": 6267, "Python id": 6271, "finished": true, "Ev Idx": 9171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907858.468, "dur": 4.670,
    "args": {
      "Python parent id": 6267, "Python id": 6272, "finished": true, "Ev Idx": 9172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907858.663, "dur": 4.420,
    "args": {
      "Python parent id": 6272, "Python id": 6273, "finished": true, "Ev Idx": 9173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907858.792, "dur": 2.696,
    "args": {
      "Python parent id": 6273, "Python id": 6274, "finished": true, "Ev Idx": 9174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907858.998, "dur": 2.447,
    "args": {
      "Python parent id": 6274, "Python id": 6275, "finished": true, "Ev Idx": 9175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.115, "dur": 0.206,
    "args": {
      "Python parent id": 6275, "Python id": 6276, "finished": true, "Ev Idx": 9176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.221, "dur": 0.057,
    "args": {
      "Python parent id": 6276, "Python id": 6277, "finished": true, "Ev Idx": 9177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.503, "dur": 1.667,
    "args": {
      "Python parent id": 6275, "Python id": 6278, "finished": true, "Ev Idx": 9178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.685, "dur": 1.427,
    "args": {
      "Python parent id": 6278, "Python id": 6279, "finished": true, "Ev Idx": 9179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.850, "dur": 1.218,
    "args": {
      "Python parent id": 6279, "Python id": 6280, "finished": true, "Ev Idx": 9180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907859.966, "dur": 0.942,
    "args": {
      "Python parent id": 6280, "Python id": 6281, "finished": true, "Ev Idx": 9181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907860.103, "dur": 0.435,
    "args": {
      "Python parent id": 6281, "Python id": 6282, "finished": true, "Ev Idx": 9182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907860.186, "dur": 0.044,
    "args": {
      "Python parent id": 6282, "Python id": 6283, "finished": true, "Ev Idx": 9183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907860.367, "dur": 0.128,
    "args": {
      "Python parent id": 6282, "Python id": 6284, "finished": true, "Ev Idx": 9184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907861.345, "dur": 0.064,
    "args": {
      "Python parent id": 6275, "Python id": 6285, "finished": true, "Ev Idx": 9185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907861.637, "dur": 0.076,
    "args": {
      "Python parent id": 6273, "Python id": 6286, "finished": true, "Ev Idx": 9186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907861.991, "dur": 1.052,
    "args": {
      "Python parent id": 6273, "Python id": 6287, "finished": true, "Ev Idx": 9187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907862.125, "dur": 0.875,
    "args": {
      "Python parent id": 6287, "Python id": 6288, "finished": true, "Ev Idx": 9188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907862.257, "dur": 0.355,
    "args": {
      "Python parent id": 6288, "Python id": 6289, "finished": true, "Ev Idx": 9189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907862.384, "dur": 0.172,
    "args": {
      "Python parent id": 6289, "Python id": 6290, "finished": true, "Ev Idx": 9190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907862.467, "dur": 0.060,
    "args": {
      "Python parent id": 6290, "Python id": 6291, "finished": true, "Ev Idx": 9191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907862.736, "dur": 0.231,
    "args": {
      "Python parent id": 6288, "Python id": 6292, "finished": true, "Ev Idx": 9192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907863.393, "dur": 0.148,
    "args": {
      "Python parent id": 6257, "Python id": 6293, "finished": true, "Ev Idx": 9193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907864.060, "dur": 0.849,
    "args": {
      "Python parent id": 6257, "Python id": 6294, "finished": true, "Ev Idx": 9194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907864.451, "dur": 0.406,
    "args": {
      "Python parent id": 6294, "Python id": 6295, "finished": true, "Ev Idx": 9195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907865.301, "dur": 2.983,
    "args": {
      "Python parent id": 6256, "Python id": 6296, "finished": true, "Ev Idx": 9196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740907869.216, "dur": 86.308,
    "args": {
      "Python parent id": 12, "Python id": 6297, "finished": true, "Ev Idx": 9197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907869.502, "dur": 0.491,
    "args": {
      "Python parent id": 6297, "Python id": 6298, "finished": true, "Ev Idx": 9198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907869.788, "dur": 0.085,
    "args": {
      "Python parent id": 6298, "Python id": 6299, "finished": true, "Ev Idx": 9199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907870.224, "dur": 0.327,
    "args": {
      "Python parent id": 6297, "Python id": 6300, "finished": true, "Ev Idx": 9200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740907870.764, "dur": 0.281,
    "args": {
      "Python parent id": 6297, "Python id": 6301, "finished": true, "Ev Idx": 9201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740907871.272, "dur": 1.504,
    "args": {
      "Python parent id": 6297, "Python id": 6302, "finished": true, "Ev Idx": 9202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740907873.334, "dur": 82.075,
    "args": {
      "Python parent id": 6297, "Python id": 6303, "finished": true, "Ev Idx": 9203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740907874.058, "dur": 81.177,
    "args": {
      "Python parent id": 6303, "Python id": 6304, "finished": true, "Ev Idx": 9204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907874.242, "dur": 80.932,
    "args": {
      "Python parent id": 6304, "Python id": 6305, "finished": true, "Ev Idx": 9205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740907878.957, "dur": 75.043,
    "args": {
      "Python parent id": 6305, "Python id": 6306, "finished": true, "Ev Idx": 9206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740907879.246, "dur": 74.682,
    "args": {
      "Python parent id": 6306, "Python id": 6307, "finished": true, "Ev Idx": 9207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740907879.777, "dur": 74.019,
    "args": {
      "Python parent id": 6307, "Python id": 6308, "finished": true, "Ev Idx": 9208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907880.215, "dur": 0.325,
    "args": {
      "Python parent id": 6308, "Python id": 6309, "finished": true, "Ev Idx": 9209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907880.447, "dur": 0.038,
    "args": {
      "Python parent id": 6309, "Python id": 6310, "finished": true, "Ev Idx": 9210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740907880.849, "dur": 0.365,
    "args": {
      "Python parent id": 6308, "Python id": 6311, "finished": true, "Ev Idx": 9211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740907881.110, "dur": 0.065,
    "args": {
      "Python parent id": 6311, "Python id": 6312, "finished": true, "Ev Idx": 9212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907881.556, "dur": 72.139,
    "args": {
      "Python parent id": 6308, "Python id": 6313, "finished": true, "Ev Idx": 9213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740907882.230, "dur": 70.306,
    "args": {
      "Python parent id": 6313, "Python id": 6314, "finished": true, "Ev Idx": 9214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740907888.627, "dur": 15.942,
    "args": {
      "Python parent id": 6314, "Python id": 6315, "finished": true, "Ev Idx": 9215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740907888.743, "dur": 0.122,
    "args": {
      "Python parent id": 6315, "Python id": 6316, "finished": true, "Ev Idx": 9216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740907889.102, "dur": 15.385,
    "args": {
      "Python parent id": 6315, "Python id": 6317, "finished": true, "Ev Idx": 9217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907889.252, "dur": 6.788,
    "args": {
      "Python parent id": 6317, "Python id": 6318, "finished": true, "Ev Idx": 9218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907913.376, "dur": 6.726,
    "args": {
      "Python parent id": 6314, "Python id": 6319, "finished": true, "Ev Idx": 9219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740907927.630, "dur": 10.533,
    "args": {
      "Python parent id": 6314, "Python id": 6320, "finished": true, "Ev Idx": 9220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907956.623, "dur": 15.773,
    "args": {
      "Python parent id": 12, "Python id": 6321, "finished": true, "Ev Idx": 9221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907956.982, "dur": 12.210,
    "args": {
      "Python parent id": 6321, "Python id": 6322, "finished": true, "Ev Idx": 9222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907957.195, "dur": 0.508,
    "args": {
      "Python parent id": 6322, "Python id": 6323, "finished": true, "Ev Idx": 9223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907957.396, "dur": 0.223,
    "args": {
      "Python parent id": 6323, "Python id": 6324, "finished": true, "Ev Idx": 9224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907957.507, "dur": 0.083,
    "args": {
      "Python parent id": 6324, "Python id": 6325, "finished": true, "Ev Idx": 9225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907957.999, "dur": 9.014,
    "args": {
      "Python parent id": 6322, "Python id": 6326, "finished": true, "Ev Idx": 9226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907958.146, "dur": 0.116,
    "args": {
      "Python parent id": 6326, "Python id": 6327, "finished": true, "Ev Idx": 9227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907958.386, "dur": 0.056,
    "args": {
      "Python parent id": 6326, "Python id": 6328, "finished": true, "Ev Idx": 9228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907958.608, "dur": 0.052,
    "args": {
      "Python parent id": 6326, "Python id": 6329, "finished": true, "Ev Idx": 9229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907959.049, "dur": 0.038,
    "args": {
      "Python parent id": 6326, "Python id": 6330, "finished": true, "Ev Idx": 9230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907959.361, "dur": 0.071,
    "args": {
      "Python parent id": 6326, "Python id": 6331, "finished": true, "Ev Idx": 9231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907959.798, "dur": 7.132,
    "args": {
      "Python parent id": 6326, "Python id": 6332, "finished": true, "Ev Idx": 9232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907959.903, "dur": 0.062,
    "args": {
      "Python parent id": 6332, "Python id": 6333, "finished": true, "Ev Idx": 9233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907960.113, "dur": 0.059,
    "args": {
      "Python parent id": 6332, "Python id": 6334, "finished": true, "Ev Idx": 9234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907960.352, "dur": 0.071,
    "args": {
      "Python parent id": 6332, "Python id": 6335, "finished": true, "Ev Idx": 9235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907960.714, "dur": 0.052,
    "args": {
      "Python parent id": 6332, "Python id": 6336, "finished": true, "Ev Idx": 9236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.086, "dur": 5.744,
    "args": {
      "Python parent id": 6332, "Python id": 6337, "finished": true, "Ev Idx": 9237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.355, "dur": 5.416,
    "args": {
      "Python parent id": 6337, "Python id": 6338, "finished": true, "Ev Idx": 9238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.486, "dur": 3.500,
    "args": {
      "Python parent id": 6338, "Python id": 6339, "finished": true, "Ev Idx": 9239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.691, "dur": 3.254,
    "args": {
      "Python parent id": 6339, "Python id": 6340, "finished": true, "Ev Idx": 9240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.811, "dur": 0.299,
    "args": {
      "Python parent id": 6340, "Python id": 6341, "finished": true, "Ev Idx": 9241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907961.936, "dur": 0.093,
    "args": {
      "Python parent id": 6341, "Python id": 6342, "finished": true, "Ev Idx": 9242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907962.315, "dur": 2.251,
    "args": {
      "Python parent id": 6340, "Python id": 6343, "finished": true, "Ev Idx": 9243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907962.517, "dur": 1.962,
    "args": {
      "Python parent id": 6343, "Python id": 6344, "finished": true, "Ev Idx": 9244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907962.759, "dur": 1.654,
    "args": {
      "Python parent id": 6344, "Python id": 6345, "finished": true, "Ev Idx": 9245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907962.931, "dur": 1.220,
    "args": {
      "Python parent id": 6345, "Python id": 6346, "finished": true, "Ev Idx": 9246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907963.125, "dur": 0.532,
    "args": {
      "Python parent id": 6346, "Python id": 6347, "finished": true, "Ev Idx": 9247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907963.226, "dur": 0.043,
    "args": {
      "Python parent id": 6347, "Python id": 6348, "finished": true, "Ev Idx": 9248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907963.435, "dur": 0.166,
    "args": {
      "Python parent id": 6347, "Python id": 6349, "finished": true, "Ev Idx": 9249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740907964.730, "dur": 0.156,
    "args": {
      "Python parent id": 6340, "Python id": 6350, "finished": true, "Ev Idx": 9250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740907965.141, "dur": 0.100,
    "args": {
      "Python parent id": 6338, "Python id": 6351, "finished": true, "Ev Idx": 9251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740907965.525, "dur": 1.202,
    "args": {
      "Python parent id": 6338, "Python id": 6352, "finished": true, "Ev Idx": 9252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740907965.668, "dur": 1.017,
    "args": {
      "Python parent id": 6352, "Python id": 6353, "finished": true, "Ev Idx": 9253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907965.812, "dur": 0.371,
    "args": {
      "Python parent id": 6353, "Python id": 6354, "finished": true, "Ev Idx": 9254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907965.957, "dur": 0.171,
    "args": {
      "Python parent id": 6354, "Python id": 6355, "finished": true, "Ev Idx": 9255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907966.047, "dur": 0.054,
    "args": {
      "Python parent id": 6355, "Python id": 6356, "finished": true, "Ev Idx": 9256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740907966.301, "dur": 0.355,
    "args": {
      "Python parent id": 6353, "Python id": 6357, "finished": true, "Ev Idx": 9257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740907967.158, "dur": 0.260,
    "args": {
      "Python parent id": 6322, "Python id": 6358, "finished": true, "Ev Idx": 9258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740907967.988, "dur": 1.068,
    "args": {
      "Python parent id": 6322, "Python id": 6359, "finished": true, "Ev Idx": 9259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740907968.502, "dur": 0.509,
    "args": {
      "Python parent id": 6359, "Python id": 6360, "finished": true, "Ev Idx": 9260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740907969.466, "dur": 2.858,
    "args": {
      "Python parent id": 6321, "Python id": 6361, "finished": true, "Ev Idx": 9261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740907972.897, "dur": 7.813,
    "args": {
      "Python parent id": 12, "Python id": 6362, "finished": true, "Ev Idx": 9262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740907981.192, "dur": 0.088,
    "args": {
      "Python parent id": 12, "Python id": 6363, "finished": true, "Ev Idx": 9263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907981.960, "dur": 0.182,
    "args": {
      "Python parent id": 12, "Python id": 6364, "finished": true, "Ev Idx": 9264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907982.606, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 6365, "finished": true, "Ev Idx": 9265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740907983.037, "dur": 9.307,
    "args": {
      "Python parent id": 12, "Python id": 6366, "finished": true, "Ev Idx": 9266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740907993.519, "dur": 15.173,
    "args": {
      "Python parent id": 12, "Python id": 6367, "finished": true, "Ev Idx": 9267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740907993.825, "dur": 10.483,
    "args": {
      "Python parent id": 6367, "Python id": 6368, "finished": true, "Ev Idx": 9268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740907993.981, "dur": 0.537,
    "args": {
      "Python parent id": 6368, "Python id": 6369, "finished": true, "Ev Idx": 9269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740907994.176, "dur": 0.247,
    "args": {
      "Python parent id": 6369, "Python id": 6370, "finished": true, "Ev Idx": 9270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740907994.286, "dur": 0.106,
    "args": {
      "Python parent id": 6370, "Python id": 6371, "finished": true, "Ev Idx": 9271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907994.774, "dur": 7.756,
    "args": {
      "Python parent id": 6368, "Python id": 6372, "finished": true, "Ev Idx": 9272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907994.884, "dur": 0.065,
    "args": {
      "Python parent id": 6372, "Python id": 6373, "finished": true, "Ev Idx": 9273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907995.057, "dur": 0.059,
    "args": {
      "Python parent id": 6372, "Python id": 6374, "finished": true, "Ev Idx": 9274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907995.273, "dur": 0.061,
    "args": {
      "Python parent id": 6372, "Python id": 6375, "finished": true, "Ev Idx": 9275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907995.720, "dur": 0.060,
    "args": {
      "Python parent id": 6372, "Python id": 6376, "finished": true, "Ev Idx": 9276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907996.012, "dur": 0.073,
    "args": {
      "Python parent id": 6372, "Python id": 6377, "finished": true, "Ev Idx": 9277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907996.369, "dur": 6.069,
    "args": {
      "Python parent id": 6372, "Python id": 6378, "finished": true, "Ev Idx": 9278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907996.467, "dur": 0.062,
    "args": {
      "Python parent id": 6378, "Python id": 6379, "finished": true, "Ev Idx": 9279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907996.675, "dur": 0.057,
    "args": {
      "Python parent id": 6378, "Python id": 6380, "finished": true, "Ev Idx": 9280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907996.878, "dur": 0.073,
    "args": {
      "Python parent id": 6378, "Python id": 6381, "finished": true, "Ev Idx": 9281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740907997.252, "dur": 0.035,
    "args": {
      "Python parent id": 6378, "Python id": 6382, "finished": true, "Ev Idx": 9282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740907997.592, "dur": 4.795,
    "args": {
      "Python parent id": 6378, "Python id": 6383, "finished": true, "Ev Idx": 9283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740907997.837, "dur": 4.496,
    "args": {
      "Python parent id": 6383, "Python id": 6384, "finished": true, "Ev Idx": 9284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740907997.954, "dur": 2.834,
    "args": {
      "Python parent id": 6384, "Python id": 6385, "finished": true, "Ev Idx": 9285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740907998.124, "dur": 2.624,
    "args": {
      "Python parent id": 6385, "Python id": 6386, "finished": true, "Ev Idx": 9286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740907998.245, "dur": 0.259,
    "args": {
      "Python parent id": 6386, "Python id": 6387, "finished": true, "Ev Idx": 9287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740907998.373, "dur": 0.055,
    "args": {
      "Python parent id": 6387, "Python id": 6388, "finished": true, "Ev Idx": 9288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740907998.664, "dur": 1.779,
    "args": {
      "Python parent id": 6386, "Python id": 6389, "finished": true, "Ev Idx": 9289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740907998.830, "dur": 1.555,
    "args": {
      "Python parent id": 6389, "Python id": 6390, "finished": true, "Ev Idx": 9290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740907999.024, "dur": 1.316,
    "args": {
      "Python parent id": 6390, "Python id": 6391, "finished": true, "Ev Idx": 9291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740907999.162, "dur": 1.008,
    "args": {
      "Python parent id": 6391, "Python id": 6392, "finished": true, "Ev Idx": 9292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740907999.359, "dur": 0.421,
    "args": {
      "Python parent id": 6392, "Python id": 6393, "finished": true, "Ev Idx": 9293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740907999.437, "dur": 0.041,
    "args": {
      "Python parent id": 6393, "Python id": 6394, "finished": true, "Ev Idx": 9294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740907999.626, "dur": 0.096,
    "args": {
      "Python parent id": 6393, "Python id": 6395, "finished": true, "Ev Idx": 9295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908000.610, "dur": 0.090,
    "args": {
      "Python parent id": 6386, "Python id": 6396, "finished": true, "Ev Idx": 9296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908000.928, "dur": 0.092,
    "args": {
      "Python parent id": 6384, "Python id": 6397, "finished": true, "Ev Idx": 9297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908001.244, "dur": 1.049,
    "args": {
      "Python parent id": 6384, "Python id": 6398, "finished": true, "Ev Idx": 9298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908001.392, "dur": 0.861,
    "args": {
      "Python parent id": 6398, "Python id": 6399, "finished": true, "Ev Idx": 9299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908001.524, "dur": 0.370,
    "args": {
      "Python parent id": 6399, "Python id": 6400, "finished": true, "Ev Idx": 9300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908001.650, "dur": 0.185,
    "args": {
      "Python parent id": 6400, "Python id": 6401, "finished": true, "Ev Idx": 9301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908001.759, "dur": 0.048,
    "args": {
      "Python parent id": 6401, "Python id": 6402, "finished": true, "Ev Idx": 9302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908002.011, "dur": 0.216,
    "args": {
      "Python parent id": 6399, "Python id": 6403, "finished": true, "Ev Idx": 9303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908002.633, "dur": 0.202,
    "args": {
      "Python parent id": 6368, "Python id": 6404, "finished": true, "Ev Idx": 9304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908003.366, "dur": 0.827,
    "args": {
      "Python parent id": 6368, "Python id": 6405, "finished": true, "Ev Idx": 9305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908003.729, "dur": 0.418,
    "args": {
      "Python parent id": 6405, "Python id": 6406, "finished": true, "Ev Idx": 9306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908004.582, "dur": 4.053,
    "args": {
      "Python parent id": 6367, "Python id": 6407, "finished": true, "Ev Idx": 9307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908009.520, "dur": 87.990,
    "args": {
      "Python parent id": 12, "Python id": 6408, "finished": true, "Ev Idx": 9308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908009.860, "dur": 0.437,
    "args": {
      "Python parent id": 6408, "Python id": 6409, "finished": true, "Ev Idx": 9309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908010.135, "dur": 0.101,
    "args": {
      "Python parent id": 6409, "Python id": 6410, "finished": true, "Ev Idx": 9310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908010.611, "dur": 0.322,
    "args": {
      "Python parent id": 6408, "Python id": 6411, "finished": true, "Ev Idx": 9311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908011.118, "dur": 0.284,
    "args": {
      "Python parent id": 6408, "Python id": 6412, "finished": true, "Ev Idx": 9312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908011.624, "dur": 1.535,
    "args": {
      "Python parent id": 6408, "Python id": 6413, "finished": true, "Ev Idx": 9313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908013.717, "dur": 83.692,
    "args": {
      "Python parent id": 6408, "Python id": 6414, "finished": true, "Ev Idx": 9314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908014.468, "dur": 82.779,
    "args": {
      "Python parent id": 6414, "Python id": 6415, "finished": true, "Ev Idx": 9315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908014.708, "dur": 82.467,
    "args": {
      "Python parent id": 6415, "Python id": 6416, "finished": true, "Ev Idx": 9316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908019.342, "dur": 76.611,
    "args": {
      "Python parent id": 6416, "Python id": 6417, "finished": true, "Ev Idx": 9317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908019.658, "dur": 76.243,
    "args": {
      "Python parent id": 6417, "Python id": 6418, "finished": true, "Ev Idx": 9318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908020.139, "dur": 75.639,
    "args": {
      "Python parent id": 6418, "Python id": 6419, "finished": true, "Ev Idx": 9319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908020.602, "dur": 0.402,
    "args": {
      "Python parent id": 6419, "Python id": 6420, "finished": true, "Ev Idx": 9320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908020.900, "dur": 0.055,
    "args": {
      "Python parent id": 6420, "Python id": 6421, "finished": true, "Ev Idx": 9321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908021.303, "dur": 0.324,
    "args": {
      "Python parent id": 6419, "Python id": 6422, "finished": true, "Ev Idx": 9322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908021.538, "dur": 0.040,
    "args": {
      "Python parent id": 6422, "Python id": 6423, "finished": true, "Ev Idx": 9323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908021.971, "dur": 73.714,
    "args": {
      "Python parent id": 6419, "Python id": 6424, "finished": true, "Ev Idx": 9324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908022.613, "dur": 71.967,
    "args": {
      "Python parent id": 6424, "Python id": 6425, "finished": true, "Ev Idx": 9325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908029.312, "dur": 17.913,
    "args": {
      "Python parent id": 6425, "Python id": 6426, "finished": true, "Ev Idx": 9326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908029.442, "dur": 0.116,
    "args": {
      "Python parent id": 6426, "Python id": 6427, "finished": true, "Ev Idx": 9327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908029.809, "dur": 17.348,
    "args": {
      "Python parent id": 6426, "Python id": 6428, "finished": true, "Ev Idx": 9328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908029.928, "dur": 7.786,
    "args": {
      "Python parent id": 6428, "Python id": 6429, "finished": true, "Ev Idx": 9329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908054.809, "dur": 6.610,
    "args": {
      "Python parent id": 6425, "Python id": 6430, "finished": true, "Ev Idx": 9330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908071.009, "dur": 7.825,
    "args": {
      "Python parent id": 6425, "Python id": 6431, "finished": true, "Ev Idx": 9331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908098.623, "dur": 16.126,
    "args": {
      "Python parent id": 12, "Python id": 6432, "finished": true, "Ev Idx": 9332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908098.955, "dur": 11.653,
    "args": {
      "Python parent id": 6432, "Python id": 6433, "finished": true, "Ev Idx": 9333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908099.145, "dur": 0.408,
    "args": {
      "Python parent id": 6433, "Python id": 6434, "finished": true, "Ev Idx": 9334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908099.279, "dur": 0.215,
    "args": {
      "Python parent id": 6434, "Python id": 6435, "finished": true, "Ev Idx": 9335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908099.368, "dur": 0.092,
    "args": {
      "Python parent id": 6435, "Python id": 6436, "finished": true, "Ev Idx": 9336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908099.860, "dur": 8.623,
    "args": {
      "Python parent id": 6433, "Python id": 6437, "finished": true, "Ev Idx": 9337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908099.975, "dur": 0.087,
    "args": {
      "Python parent id": 6437, "Python id": 6438, "finished": true, "Ev Idx": 9338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908100.180, "dur": 0.058,
    "args": {
      "Python parent id": 6437, "Python id": 6439, "finished": true, "Ev Idx": 9339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908100.376, "dur": 0.054,
    "args": {
      "Python parent id": 6437, "Python id": 6440, "finished": true, "Ev Idx": 9340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908100.865, "dur": 0.051,
    "args": {
      "Python parent id": 6437, "Python id": 6441, "finished": true, "Ev Idx": 9341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908101.215, "dur": 0.075,
    "args": {
      "Python parent id": 6437, "Python id": 6442, "finished": true, "Ev Idx": 9342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908101.646, "dur": 6.760,
    "args": {
      "Python parent id": 6437, "Python id": 6443, "finished": true, "Ev Idx": 9343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908101.765, "dur": 0.061,
    "args": {
      "Python parent id": 6443, "Python id": 6444, "finished": true, "Ev Idx": 9344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908101.975, "dur": 0.058,
    "args": {
      "Python parent id": 6443, "Python id": 6445, "finished": true, "Ev Idx": 9345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908102.224, "dur": 0.072,
    "args": {
      "Python parent id": 6443, "Python id": 6446, "finished": true, "Ev Idx": 9346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908102.624, "dur": 0.049,
    "args": {
      "Python parent id": 6443, "Python id": 6447, "finished": true, "Ev Idx": 9347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.009, "dur": 5.314,
    "args": {
      "Python parent id": 6443, "Python id": 6448, "finished": true, "Ev Idx": 9348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.255, "dur": 5.012,
    "args": {
      "Python parent id": 6448, "Python id": 6449, "finished": true, "Ev Idx": 9349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.386, "dur": 3.162,
    "args": {
      "Python parent id": 6449, "Python id": 6450, "finished": true, "Ev Idx": 9350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.594, "dur": 2.911,
    "args": {
      "Python parent id": 6450, "Python id": 6451, "finished": true, "Ev Idx": 9351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.752, "dur": 0.269,
    "args": {
      "Python parent id": 6451, "Python id": 6452, "finished": true, "Ev Idx": 9352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908103.890, "dur": 0.061,
    "args": {
      "Python parent id": 6452, "Python id": 6453, "finished": true, "Ev Idx": 9353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.179, "dur": 1.957,
    "args": {
      "Python parent id": 6451, "Python id": 6454, "finished": true, "Ev Idx": 9354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.372, "dur": 1.670,
    "args": {
      "Python parent id": 6454, "Python id": 6455, "finished": true, "Ev Idx": 9355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.544, "dur": 1.452,
    "args": {
      "Python parent id": 6455, "Python id": 6456, "finished": true, "Ev Idx": 9356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.687, "dur": 1.076,
    "args": {
      "Python parent id": 6456, "Python id": 6457, "finished": true, "Ev Idx": 9357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.859, "dur": 0.486,
    "args": {
      "Python parent id": 6457, "Python id": 6458, "finished": true, "Ev Idx": 9358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908104.954, "dur": 0.046,
    "args": {
      "Python parent id": 6458, "Python id": 6459, "finished": true, "Ev Idx": 9359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908105.157, "dur": 0.121,
    "args": {
      "Python parent id": 6458, "Python id": 6460, "finished": true, "Ev Idx": 9360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908106.323, "dur": 0.147,
    "args": {
      "Python parent id": 6451, "Python id": 6461, "finished": true, "Ev Idx": 9361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908106.719, "dur": 0.119,
    "args": {
      "Python parent id": 6449, "Python id": 6462, "finished": true, "Ev Idx": 9362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.085, "dur": 1.138,
    "args": {
      "Python parent id": 6449, "Python id": 6463, "finished": true, "Ev Idx": 9363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.241, "dur": 0.940,
    "args": {
      "Python parent id": 6463, "Python id": 6464, "finished": true, "Ev Idx": 9364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.369, "dur": 0.353,
    "args": {
      "Python parent id": 6464, "Python id": 6465, "finished": true, "Ev Idx": 9365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.499, "dur": 0.169,
    "args": {
      "Python parent id": 6465, "Python id": 6466, "finished": true, "Ev Idx": 9366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.582, "dur": 0.058,
    "args": {
      "Python parent id": 6466, "Python id": 6467, "finished": true, "Ev Idx": 9367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908107.842, "dur": 0.308,
    "args": {
      "Python parent id": 6464, "Python id": 6468, "finished": true, "Ev Idx": 9368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908108.639, "dur": 0.244,
    "args": {
      "Python parent id": 6433, "Python id": 6469, "finished": true, "Ev Idx": 9369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908109.478, "dur": 0.996,
    "args": {
      "Python parent id": 6433, "Python id": 6470, "finished": true, "Ev Idx": 9370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908109.880, "dur": 0.547,
    "args": {
      "Python parent id": 6470, "Python id": 6471, "finished": true, "Ev Idx": 9371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908110.937, "dur": 3.742,
    "args": {
      "Python parent id": 6432, "Python id": 6472, "finished": true, "Ev Idx": 9372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908115.271, "dur": 7.706,
    "args": {
      "Python parent id": 12, "Python id": 6473, "finished": true, "Ev Idx": 9373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908123.460, "dur": 0.136,
    "args": {
      "Python parent id": 12, "Python id": 6474, "finished": true, "Ev Idx": 9374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908124.200, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 6475, "finished": true, "Ev Idx": 9375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908124.869, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 6476, "finished": true, "Ev Idx": 9376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908125.336, "dur": 10.190,
    "args": {
      "Python parent id": 12, "Python id": 6477, "finished": true, "Ev Idx": 9377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908136.750, "dur": 15.257,
    "args": {
      "Python parent id": 12, "Python id": 6478, "finished": true, "Ev Idx": 9378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908137.072, "dur": 10.706,
    "args": {
      "Python parent id": 6478, "Python id": 6479, "finished": true, "Ev Idx": 9379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908137.216, "dur": 0.456,
    "args": {
      "Python parent id": 6479, "Python id": 6480, "finished": true, "Ev Idx": 9380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908137.365, "dur": 0.229,
    "args": {
      "Python parent id": 6480, "Python id": 6481, "finished": true, "Ev Idx": 9381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908137.464, "dur": 0.095,
    "args": {
      "Python parent id": 6481, "Python id": 6482, "finished": true, "Ev Idx": 9382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908137.924, "dur": 7.950,
    "args": {
      "Python parent id": 6479, "Python id": 6483, "finished": true, "Ev Idx": 9383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908138.040, "dur": 0.061,
    "args": {
      "Python parent id": 6483, "Python id": 6484, "finished": true, "Ev Idx": 9384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908138.205, "dur": 0.057,
    "args": {
      "Python parent id": 6483, "Python id": 6485, "finished": true, "Ev Idx": 9385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908138.399, "dur": 0.052,
    "args": {
      "Python parent id": 6483, "Python id": 6486, "finished": true, "Ev Idx": 9386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908138.844, "dur": 0.034,
    "args": {
      "Python parent id": 6483, "Python id": 6487, "finished": true, "Ev Idx": 9387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908139.128, "dur": 0.072,
    "args": {
      "Python parent id": 6483, "Python id": 6488, "finished": true, "Ev Idx": 9388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908139.512, "dur": 6.279,
    "args": {
      "Python parent id": 6483, "Python id": 6489, "finished": true, "Ev Idx": 9389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908139.586, "dur": 0.060,
    "args": {
      "Python parent id": 6489, "Python id": 6490, "finished": true, "Ev Idx": 9390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908139.781, "dur": 0.056,
    "args": {
      "Python parent id": 6489, "Python id": 6491, "finished": true, "Ev Idx": 9391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908140.003, "dur": 0.059,
    "args": {
      "Python parent id": 6489, "Python id": 6492, "finished": true, "Ev Idx": 9392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908140.402, "dur": 0.039,
    "args": {
      "Python parent id": 6489, "Python id": 6493, "finished": true, "Ev Idx": 9393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908140.751, "dur": 4.958,
    "args": {
      "Python parent id": 6489, "Python id": 6494, "finished": true, "Ev Idx": 9394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908140.921, "dur": 4.731,
    "args": {
      "Python parent id": 6494, "Python id": 6495, "finished": true, "Ev Idx": 9395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908141.132, "dur": 2.894,
    "args": {
      "Python parent id": 6495, "Python id": 6496, "finished": true, "Ev Idx": 9396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908141.333, "dur": 2.649,
    "args": {
      "Python parent id": 6496, "Python id": 6497, "finished": true, "Ev Idx": 9397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908141.467, "dur": 0.292,
    "args": {
      "Python parent id": 6497, "Python id": 6498, "finished": true, "Ev Idx": 9398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908141.610, "dur": 0.108,
    "args": {
      "Python parent id": 6498, "Python id": 6499, "finished": true, "Ev Idx": 9399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908141.901, "dur": 1.804,
    "args": {
      "Python parent id": 6497, "Python id": 6500, "finished": true, "Ev Idx": 9400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.103, "dur": 1.545,
    "args": {
      "Python parent id": 6500, "Python id": 6501, "finished": true, "Ev Idx": 9401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.251, "dur": 1.353,
    "args": {
      "Python parent id": 6501, "Python id": 6502, "finished": true, "Ev Idx": 9402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.381, "dur": 1.029,
    "args": {
      "Python parent id": 6502, "Python id": 6503, "finished": true, "Ev Idx": 9403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.537, "dur": 0.433,
    "args": {
      "Python parent id": 6503, "Python id": 6504, "finished": true, "Ev Idx": 9404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.612, "dur": 0.043,
    "args": {
      "Python parent id": 6504, "Python id": 6505, "finished": true, "Ev Idx": 9405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908142.812, "dur": 0.118,
    "args": {
      "Python parent id": 6504, "Python id": 6506, "finished": true, "Ev Idx": 9406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908143.882, "dur": 0.065,
    "args": {
      "Python parent id": 6497, "Python id": 6507, "finished": true, "Ev Idx": 9407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.172, "dur": 0.075,
    "args": {
      "Python parent id": 6495, "Python id": 6508, "finished": true, "Ev Idx": 9408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.478, "dur": 1.131,
    "args": {
      "Python parent id": 6495, "Python id": 6509, "finished": true, "Ev Idx": 9409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.626, "dur": 0.838,
    "args": {
      "Python parent id": 6509, "Python id": 6510, "finished": true, "Ev Idx": 9410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.757, "dur": 0.367,
    "args": {
      "Python parent id": 6510, "Python id": 6511, "finished": true, "Ev Idx": 9411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.894, "dur": 0.176,
    "args": {
      "Python parent id": 6511, "Python id": 6512, "finished": true, "Ev Idx": 9412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908144.982, "dur": 0.052,
    "args": {
      "Python parent id": 6512, "Python id": 6513, "finished": true, "Ev Idx": 9413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908145.248, "dur": 0.187,
    "args": {
      "Python parent id": 6510, "Python id": 6514, "finished": true, "Ev Idx": 9414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908146.113, "dur": 0.131,
    "args": {
      "Python parent id": 6479, "Python id": 6515, "finished": true, "Ev Idx": 9415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908146.792, "dur": 0.873,
    "args": {
      "Python parent id": 6479, "Python id": 6516, "finished": true, "Ev Idx": 9416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908147.187, "dur": 0.431,
    "args": {
      "Python parent id": 6516, "Python id": 6517, "finished": true, "Ev Idx": 9417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908148.038, "dur": 3.915,
    "args": {
      "Python parent id": 6478, "Python id": 6518, "finished": true, "Ev Idx": 9418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908152.887, "dur": 91.587,
    "args": {
      "Python parent id": 12, "Python id": 6519, "finished": true, "Ev Idx": 9419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908153.268, "dur": 0.426,
    "args": {
      "Python parent id": 6519, "Python id": 6520, "finished": true, "Ev Idx": 9420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908153.507, "dur": 0.081,
    "args": {
      "Python parent id": 6520, "Python id": 6521, "finished": true, "Ev Idx": 9421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908153.990, "dur": 0.355,
    "args": {
      "Python parent id": 6519, "Python id": 6522, "finished": true, "Ev Idx": 9422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908154.529, "dur": 0.247,
    "args": {
      "Python parent id": 6519, "Python id": 6523, "finished": true, "Ev Idx": 9423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908155.045, "dur": 1.400,
    "args": {
      "Python parent id": 6519, "Python id": 6524, "finished": true, "Ev Idx": 9424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908157.007, "dur": 87.374,
    "args": {
      "Python parent id": 6519, "Python id": 6525, "finished": true, "Ev Idx": 9425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908157.734, "dur": 86.397,
    "args": {
      "Python parent id": 6525, "Python id": 6526, "finished": true, "Ev Idx": 9426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908157.971, "dur": 86.099,
    "args": {
      "Python parent id": 6526, "Python id": 6527, "finished": true, "Ev Idx": 9427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908162.461, "dur": 80.371,
    "args": {
      "Python parent id": 6527, "Python id": 6528, "finished": true, "Ev Idx": 9428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908162.755, "dur": 79.997,
    "args": {
      "Python parent id": 6528, "Python id": 6529, "finished": true, "Ev Idx": 9429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908163.284, "dur": 79.358,
    "args": {
      "Python parent id": 6529, "Python id": 6530, "finished": true, "Ev Idx": 9430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908163.761, "dur": 0.354,
    "args": {
      "Python parent id": 6530, "Python id": 6531, "finished": true, "Ev Idx": 9431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908163.991, "dur": 0.067,
    "args": {
      "Python parent id": 6531, "Python id": 6532, "finished": true, "Ev Idx": 9432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908164.367, "dur": 0.294,
    "args": {
      "Python parent id": 6530, "Python id": 6533, "finished": true, "Ev Idx": 9433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908164.569, "dur": 0.040,
    "args": {
      "Python parent id": 6533, "Python id": 6534, "finished": true, "Ev Idx": 9434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908165.009, "dur": 77.527,
    "args": {
      "Python parent id": 6530, "Python id": 6535, "finished": true, "Ev Idx": 9435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908165.590, "dur": 75.742,
    "args": {
      "Python parent id": 6535, "Python id": 6536, "finished": true, "Ev Idx": 9436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908172.290, "dur": 16.670,
    "args": {
      "Python parent id": 6536, "Python id": 6537, "finished": true, "Ev Idx": 9437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908172.398, "dur": 0.140,
    "args": {
      "Python parent id": 6537, "Python id": 6538, "finished": true, "Ev Idx": 9438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908172.759, "dur": 16.096,
    "args": {
      "Python parent id": 6537, "Python id": 6539, "finished": true, "Ev Idx": 9439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908172.874, "dur": 7.238,
    "args": {
      "Python parent id": 6539, "Python id": 6540, "finished": true, "Ev Idx": 9440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908196.153, "dur": 6.871,
    "args": {
      "Python parent id": 6536, "Python id": 6541, "finished": true, "Ev Idx": 9441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908211.599, "dur": 7.519,
    "args": {
      "Python parent id": 6536, "Python id": 6542, "finished": true, "Ev Idx": 9442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908246.021, "dur": 16.614,
    "args": {
      "Python parent id": 12, "Python id": 6543, "finished": true, "Ev Idx": 9443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908246.380, "dur": 11.185,
    "args": {
      "Python parent id": 6543, "Python id": 6544, "finished": true, "Ev Idx": 9444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908246.602, "dur": 0.403,
    "args": {
      "Python parent id": 6544, "Python id": 6545, "finished": true, "Ev Idx": 9445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908246.727, "dur": 0.215,
    "args": {
      "Python parent id": 6545, "Python id": 6546, "finished": true, "Ev Idx": 9446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908246.819, "dur": 0.088,
    "args": {
      "Python parent id": 6546, "Python id": 6547, "finished": true, "Ev Idx": 9447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908247.284, "dur": 8.350,
    "args": {
      "Python parent id": 6544, "Python id": 6548, "finished": true, "Ev Idx": 9448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908247.362, "dur": 0.066,
    "args": {
      "Python parent id": 6548, "Python id": 6549, "finished": true, "Ev Idx": 9449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908247.538, "dur": 0.059,
    "args": {
      "Python parent id": 6548, "Python id": 6550, "finished": true, "Ev Idx": 9450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908247.732, "dur": 0.050,
    "args": {
      "Python parent id": 6548, "Python id": 6551, "finished": true, "Ev Idx": 9451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908248.193, "dur": 0.074,
    "args": {
      "Python parent id": 6548, "Python id": 6552, "finished": true, "Ev Idx": 9452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908248.527, "dur": 0.062,
    "args": {
      "Python parent id": 6548, "Python id": 6553, "finished": true, "Ev Idx": 9453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908248.918, "dur": 6.621,
    "args": {
      "Python parent id": 6548, "Python id": 6554, "finished": true, "Ev Idx": 9454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908249.011, "dur": 0.059,
    "args": {
      "Python parent id": 6554, "Python id": 6555, "finished": true, "Ev Idx": 9455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908249.208, "dur": 0.056,
    "args": {
      "Python parent id": 6554, "Python id": 6556, "finished": true, "Ev Idx": 9456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908249.485, "dur": 0.066,
    "args": {
      "Python parent id": 6554, "Python id": 6557, "finished": true, "Ev Idx": 9457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908249.877, "dur": 0.034,
    "args": {
      "Python parent id": 6554, "Python id": 6558, "finished": true, "Ev Idx": 9458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908250.229, "dur": 5.254,
    "args": {
      "Python parent id": 6554, "Python id": 6559, "finished": true, "Ev Idx": 9459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908250.503, "dur": 4.923,
    "args": {
      "Python parent id": 6559, "Python id": 6560, "finished": true, "Ev Idx": 9460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908250.626, "dur": 3.161,
    "args": {
      "Python parent id": 6560, "Python id": 6561, "finished": true, "Ev Idx": 9461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908250.803, "dur": 2.943,
    "args": {
      "Python parent id": 6561, "Python id": 6562, "finished": true, "Ev Idx": 9462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908250.927, "dur": 0.263,
    "args": {
      "Python parent id": 6562, "Python id": 6563, "finished": true, "Ev Idx": 9463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908251.065, "dur": 0.079,
    "args": {
      "Python parent id": 6563, "Python id": 6564, "finished": true, "Ev Idx": 9464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908251.344, "dur": 2.044,
    "args": {
      "Python parent id": 6562, "Python id": 6565, "finished": true, "Ev Idx": 9465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908251.521, "dur": 1.792,
    "args": {
      "Python parent id": 6565, "Python id": 6566, "finished": true, "Ev Idx": 9466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908251.718, "dur": 1.549,
    "args": {
      "Python parent id": 6566, "Python id": 6567, "finished": true, "Ev Idx": 9467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908251.852, "dur": 1.135,
    "args": {
      "Python parent id": 6567, "Python id": 6568, "finished": true, "Ev Idx": 9468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908252.025, "dur": 0.500,
    "args": {
      "Python parent id": 6568, "Python id": 6569, "finished": true, "Ev Idx": 9469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908252.107, "dur": 0.064,
    "args": {
      "Python parent id": 6569, "Python id": 6570, "finished": true, "Ev Idx": 9470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908252.358, "dur": 0.129,
    "args": {
      "Python parent id": 6569, "Python id": 6571, "finished": true, "Ev Idx": 9471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908253.586, "dur": 0.120,
    "args": {
      "Python parent id": 6562, "Python id": 6572, "finished": true, "Ev Idx": 9472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908253.940, "dur": 0.089,
    "args": {
      "Python parent id": 6560, "Python id": 6573, "finished": true, "Ev Idx": 9473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908254.296, "dur": 1.085,
    "args": {
      "Python parent id": 6560, "Python id": 6574, "finished": true, "Ev Idx": 9474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908254.438, "dur": 0.900,
    "args": {
      "Python parent id": 6574, "Python id": 6575, "finished": true, "Ev Idx": 9475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908254.558, "dur": 0.350,
    "args": {
      "Python parent id": 6575, "Python id": 6576, "finished": true, "Ev Idx": 9476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908254.685, "dur": 0.166,
    "args": {
      "Python parent id": 6576, "Python id": 6577, "finished": true, "Ev Idx": 9477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908254.768, "dur": 0.051,
    "args": {
      "Python parent id": 6577, "Python id": 6578, "finished": true, "Ev Idx": 9478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908255.032, "dur": 0.275,
    "args": {
      "Python parent id": 6575, "Python id": 6579, "finished": true, "Ev Idx": 9479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908255.743, "dur": 0.192,
    "args": {
      "Python parent id": 6544, "Python id": 6580, "finished": true, "Ev Idx": 9480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908256.507, "dur": 0.934,
    "args": {
      "Python parent id": 6544, "Python id": 6581, "finished": true, "Ev Idx": 9481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908256.900, "dur": 0.495,
    "args": {
      "Python parent id": 6581, "Python id": 6582, "finished": true, "Ev Idx": 9482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908257.799, "dur": 4.781,
    "args": {
      "Python parent id": 6543, "Python id": 6583, "finished": true, "Ev Idx": 9483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908263.152, "dur": 8.776,
    "args": {
      "Python parent id": 12, "Python id": 6584, "finished": true, "Ev Idx": 9484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908272.392, "dur": 0.131,
    "args": {
      "Python parent id": 12, "Python id": 6585, "finished": true, "Ev Idx": 9485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908273.159, "dur": 0.238,
    "args": {
      "Python parent id": 12, "Python id": 6586, "finished": true, "Ev Idx": 9486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908273.858, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 6587, "finished": true, "Ev Idx": 9487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908274.268, "dur": 11.509,
    "args": {
      "Python parent id": 12, "Python id": 6588, "finished": true, "Ev Idx": 9488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908287.093, "dur": 15.043,
    "args": {
      "Python parent id": 12, "Python id": 6589, "finished": true, "Ev Idx": 9489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908287.409, "dur": 10.399,
    "args": {
      "Python parent id": 6589, "Python id": 6590, "finished": true, "Ev Idx": 9490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908287.543, "dur": 0.475,
    "args": {
      "Python parent id": 6590, "Python id": 6591, "finished": true, "Ev Idx": 9491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908287.682, "dur": 0.262,
    "args": {
      "Python parent id": 6591, "Python id": 6592, "finished": true, "Ev Idx": 9492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908287.775, "dur": 0.117,
    "args": {
      "Python parent id": 6592, "Python id": 6593, "finished": true, "Ev Idx": 9493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908288.259, "dur": 7.832,
    "args": {
      "Python parent id": 6590, "Python id": 6594, "finished": true, "Ev Idx": 9494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908288.338, "dur": 0.067,
    "args": {
      "Python parent id": 6594, "Python id": 6595, "finished": true, "Ev Idx": 9495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908288.517, "dur": 0.062,
    "args": {
      "Python parent id": 6594, "Python id": 6596, "finished": true, "Ev Idx": 9496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908288.723, "dur": 0.056,
    "args": {
      "Python parent id": 6594, "Python id": 6597, "finished": true, "Ev Idx": 9497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908289.181, "dur": 0.035,
    "args": {
      "Python parent id": 6594, "Python id": 6598, "finished": true, "Ev Idx": 9498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908289.442, "dur": 0.072,
    "args": {
      "Python parent id": 6594, "Python id": 6599, "finished": true, "Ev Idx": 9499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908289.839, "dur": 6.165,
    "args": {
      "Python parent id": 6594, "Python id": 6600, "finished": true, "Ev Idx": 9500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908289.926, "dur": 0.058,
    "args": {
      "Python parent id": 6600, "Python id": 6601, "finished": true, "Ev Idx": 9501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908290.156, "dur": 0.058,
    "args": {
      "Python parent id": 6600, "Python id": 6602, "finished": true, "Ev Idx": 9502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908290.376, "dur": 0.061,
    "args": {
      "Python parent id": 6600, "Python id": 6603, "finished": true, "Ev Idx": 9503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908290.770, "dur": 0.034,
    "args": {
      "Python parent id": 6600, "Python id": 6604, "finished": true, "Ev Idx": 9504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.113, "dur": 4.837,
    "args": {
      "Python parent id": 6600, "Python id": 6605, "finished": true, "Ev Idx": 9505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.365, "dur": 4.529,
    "args": {
      "Python parent id": 6605, "Python id": 6606, "finished": true, "Ev Idx": 9506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.482, "dur": 2.871,
    "args": {
      "Python parent id": 6606, "Python id": 6607, "finished": true, "Ev Idx": 9507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.694, "dur": 2.612,
    "args": {
      "Python parent id": 6607, "Python id": 6608, "finished": true, "Ev Idx": 9508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.823, "dur": 0.234,
    "args": {
      "Python parent id": 6608, "Python id": 6609, "finished": true, "Ev Idx": 9509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908291.923, "dur": 0.062,
    "args": {
      "Python parent id": 6609, "Python id": 6610, "finished": true, "Ev Idx": 9510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.241, "dur": 1.774,
    "args": {
      "Python parent id": 6608, "Python id": 6611, "finished": true, "Ev Idx": 9511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.392, "dur": 1.565,
    "args": {
      "Python parent id": 6611, "Python id": 6612, "finished": true, "Ev Idx": 9512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.553, "dur": 1.348,
    "args": {
      "Python parent id": 6612, "Python id": 6613, "finished": true, "Ev Idx": 9513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.667, "dur": 0.974,
    "args": {
      "Python parent id": 6613, "Python id": 6614, "finished": true, "Ev Idx": 9514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.823, "dur": 0.431,
    "args": {
      "Python parent id": 6614, "Python id": 6615, "finished": true, "Ev Idx": 9515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908292.901, "dur": 0.045,
    "args": {
      "Python parent id": 6615, "Python id": 6616, "finished": true, "Ev Idx": 9516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908293.112, "dur": 0.101,
    "args": {
      "Python parent id": 6615, "Python id": 6617, "finished": true, "Ev Idx": 9517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908294.194, "dur": 0.080,
    "args": {
      "Python parent id": 6608, "Python id": 6618, "finished": true, "Ev Idx": 9518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908294.490, "dur": 0.076,
    "args": {
      "Python parent id": 6606, "Python id": 6619, "finished": true, "Ev Idx": 9519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908294.796, "dur": 1.058,
    "args": {
      "Python parent id": 6606, "Python id": 6620, "finished": true, "Ev Idx": 9520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908294.973, "dur": 0.839,
    "args": {
      "Python parent id": 6620, "Python id": 6621, "finished": true, "Ev Idx": 9521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908295.095, "dur": 0.351,
    "args": {
      "Python parent id": 6621, "Python id": 6622, "finished": true, "Ev Idx": 9522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908295.223, "dur": 0.167,
    "args": {
      "Python parent id": 6622, "Python id": 6623, "finished": true, "Ev Idx": 9523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908295.302, "dur": 0.058,
    "args": {
      "Python parent id": 6623, "Python id": 6624, "finished": true, "Ev Idx": 9524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908295.567, "dur": 0.214,
    "args": {
      "Python parent id": 6621, "Python id": 6625, "finished": true, "Ev Idx": 9525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908296.197, "dur": 0.161,
    "args": {
      "Python parent id": 6590, "Python id": 6626, "finished": true, "Ev Idx": 9526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908296.866, "dur": 0.803,
    "args": {
      "Python parent id": 6590, "Python id": 6627, "finished": true, "Ev Idx": 9527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908297.204, "dur": 0.417,
    "args": {
      "Python parent id": 6627, "Python id": 6628, "finished": true, "Ev Idx": 9528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908298.039, "dur": 4.043,
    "args": {
      "Python parent id": 6589, "Python id": 6629, "finished": true, "Ev Idx": 9529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908302.991, "dur": 95.469,
    "args": {
      "Python parent id": 12, "Python id": 6630, "finished": true, "Ev Idx": 9530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908303.293, "dur": 0.523,
    "args": {
      "Python parent id": 6630, "Python id": 6631, "finished": true, "Ev Idx": 9531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908303.571, "dur": 0.087,
    "args": {
      "Python parent id": 6631, "Python id": 6632, "finished": true, "Ev Idx": 9532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908304.138, "dur": 0.336,
    "args": {
      "Python parent id": 6630, "Python id": 6633, "finished": true, "Ev Idx": 9533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908304.683, "dur": 0.254,
    "args": {
      "Python parent id": 6630, "Python id": 6634, "finished": true, "Ev Idx": 9534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908305.162, "dur": 2.110,
    "args": {
      "Python parent id": 6630, "Python id": 6635, "finished": true, "Ev Idx": 9535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908307.756, "dur": 90.559,
    "args": {
      "Python parent id": 6630, "Python id": 6636, "finished": true, "Ev Idx": 9536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908308.593, "dur": 89.516,
    "args": {
      "Python parent id": 6636, "Python id": 6637, "finished": true, "Ev Idx": 9537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908308.859, "dur": 89.169,
    "args": {
      "Python parent id": 6637, "Python id": 6638, "finished": true, "Ev Idx": 9538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908313.864, "dur": 82.819,
    "args": {
      "Python parent id": 6638, "Python id": 6639, "finished": true, "Ev Idx": 9539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908314.142, "dur": 82.487,
    "args": {
      "Python parent id": 6639, "Python id": 6640, "finished": true, "Ev Idx": 9540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908314.679, "dur": 81.841,
    "args": {
      "Python parent id": 6640, "Python id": 6641, "finished": true, "Ev Idx": 9541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908315.198, "dur": 0.398,
    "args": {
      "Python parent id": 6641, "Python id": 6642, "finished": true, "Ev Idx": 9542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908315.466, "dur": 0.070,
    "args": {
      "Python parent id": 6642, "Python id": 6643, "finished": true, "Ev Idx": 9543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908315.905, "dur": 0.291,
    "args": {
      "Python parent id": 6641, "Python id": 6644, "finished": true, "Ev Idx": 9544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908316.107, "dur": 0.051,
    "args": {
      "Python parent id": 6644, "Python id": 6645, "finished": true, "Ev Idx": 9545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908316.522, "dur": 79.916,
    "args": {
      "Python parent id": 6641, "Python id": 6646, "finished": true, "Ev Idx": 9546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908317.162, "dur": 77.969,
    "args": {
      "Python parent id": 6646, "Python id": 6647, "finished": true, "Ev Idx": 9547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908324.101, "dur": 19.796,
    "args": {
      "Python parent id": 6647, "Python id": 6648, "finished": true, "Ev Idx": 9548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908324.227, "dur": 0.098,
    "args": {
      "Python parent id": 6648, "Python id": 6649, "finished": true, "Ev Idx": 9549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908324.600, "dur": 19.191,
    "args": {
      "Python parent id": 6648, "Python id": 6650, "finished": true, "Ev Idx": 9550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908324.763, "dur": 7.735,
    "args": {
      "Python parent id": 6650, "Python id": 6651, "finished": true, "Ev Idx": 9551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908351.991, "dur": 7.769,
    "args": {
      "Python parent id": 6647, "Python id": 6652, "finished": true, "Ev Idx": 9552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908369.361, "dur": 7.823,
    "args": {
      "Python parent id": 6647, "Python id": 6653, "finished": true, "Ev Idx": 9553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908399.616, "dur": 17.193,
    "args": {
      "Python parent id": 12, "Python id": 6654, "finished": true, "Ev Idx": 9554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908400.013, "dur": 12.595,
    "args": {
      "Python parent id": 6654, "Python id": 6655, "finished": true, "Ev Idx": 9555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908400.234, "dur": 0.462,
    "args": {
      "Python parent id": 6655, "Python id": 6656, "finished": true, "Ev Idx": 9556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908400.383, "dur": 0.225,
    "args": {
      "Python parent id": 6656, "Python id": 6657, "finished": true, "Ev Idx": 9557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908400.485, "dur": 0.072,
    "args": {
      "Python parent id": 6657, "Python id": 6658, "finished": true, "Ev Idx": 9558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908400.945, "dur": 9.386,
    "args": {
      "Python parent id": 6655, "Python id": 6659, "finished": true, "Ev Idx": 9559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908401.024, "dur": 0.105,
    "args": {
      "Python parent id": 6659, "Python id": 6660, "finished": true, "Ev Idx": 9560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908401.222, "dur": 0.058,
    "args": {
      "Python parent id": 6659, "Python id": 6661, "finished": true, "Ev Idx": 9561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908401.415, "dur": 0.059,
    "args": {
      "Python parent id": 6659, "Python id": 6662, "finished": true, "Ev Idx": 9562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908401.899, "dur": 0.069,
    "args": {
      "Python parent id": 6659, "Python id": 6663, "finished": true, "Ev Idx": 9563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908402.209, "dur": 0.065,
    "args": {
      "Python parent id": 6659, "Python id": 6664, "finished": true, "Ev Idx": 9564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908402.614, "dur": 7.658,
    "args": {
      "Python parent id": 6659, "Python id": 6665, "finished": true, "Ev Idx": 9565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908402.712, "dur": 0.060,
    "args": {
      "Python parent id": 6665, "Python id": 6666, "finished": true, "Ev Idx": 9566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908402.914, "dur": 0.056,
    "args": {
      "Python parent id": 6665, "Python id": 6667, "finished": true, "Ev Idx": 9567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908403.139, "dur": 0.056,
    "args": {
      "Python parent id": 6665, "Python id": 6668, "finished": true, "Ev Idx": 9568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908403.474, "dur": 0.059,
    "args": {
      "Python parent id": 6665, "Python id": 6669, "finished": true, "Ev Idx": 9569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908403.810, "dur": 6.384,
    "args": {
      "Python parent id": 6665, "Python id": 6670, "finished": true, "Ev Idx": 9570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908404.071, "dur": 6.032,
    "args": {
      "Python parent id": 6670, "Python id": 6671, "finished": true, "Ev Idx": 9571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908404.207, "dur": 4.271,
    "args": {
      "Python parent id": 6671, "Python id": 6672, "finished": true, "Ev Idx": 9572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908404.381, "dur": 4.027,
    "args": {
      "Python parent id": 6672, "Python id": 6673, "finished": true, "Ev Idx": 9573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908404.497, "dur": 1.324,
    "args": {
      "Python parent id": 6673, "Python id": 6674, "finished": true, "Ev Idx": 9574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908405.681, "dur": 0.062,
    "args": {
      "Python parent id": 6674, "Python id": 6675, "finished": true, "Ev Idx": 9575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908405.952, "dur": 2.078,
    "args": {
      "Python parent id": 6673, "Python id": 6676, "finished": true, "Ev Idx": 9576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.168, "dur": 1.775,
    "args": {
      "Python parent id": 6676, "Python id": 6677, "finished": true, "Ev Idx": 9577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.333, "dur": 1.546,
    "args": {
      "Python parent id": 6677, "Python id": 6678, "finished": true, "Ev Idx": 9578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.463, "dur": 1.185,
    "args": {
      "Python parent id": 6678, "Python id": 6679, "finished": true, "Ev Idx": 9579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.652, "dur": 0.501,
    "args": {
      "Python parent id": 6679, "Python id": 6680, "finished": true, "Ev Idx": 9580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.757, "dur": 0.050,
    "args": {
      "Python parent id": 6680, "Python id": 6681, "finished": true, "Ev Idx": 9581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908406.953, "dur": 0.140,
    "args": {
      "Python parent id": 6680, "Python id": 6682, "finished": true, "Ev Idx": 9582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908408.202, "dur": 0.151,
    "args": {
      "Python parent id": 6673, "Python id": 6683, "finished": true, "Ev Idx": 9583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908408.631, "dur": 0.102,
    "args": {
      "Python parent id": 6671, "Python id": 6684, "finished": true, "Ev Idx": 9584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908408.992, "dur": 1.072,
    "args": {
      "Python parent id": 6671, "Python id": 6685, "finished": true, "Ev Idx": 9585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908409.138, "dur": 0.884,
    "args": {
      "Python parent id": 6685, "Python id": 6686, "finished": true, "Ev Idx": 9586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908409.284, "dur": 0.359,
    "args": {
      "Python parent id": 6686, "Python id": 6687, "finished": true, "Ev Idx": 9587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908409.414, "dur": 0.174,
    "args": {
      "Python parent id": 6687, "Python id": 6688, "finished": true, "Ev Idx": 9588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908409.500, "dur": 0.056,
    "args": {
      "Python parent id": 6688, "Python id": 6689, "finished": true, "Ev Idx": 9589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908409.761, "dur": 0.231,
    "args": {
      "Python parent id": 6686, "Python id": 6690, "finished": true, "Ev Idx": 9590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908410.449, "dur": 0.269,
    "args": {
      "Python parent id": 6655, "Python id": 6691, "finished": true, "Ev Idx": 9591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908411.286, "dur": 1.182,
    "args": {
      "Python parent id": 6655, "Python id": 6692, "finished": true, "Ev Idx": 9592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908411.877, "dur": 0.543,
    "args": {
      "Python parent id": 6692, "Python id": 6693, "finished": true, "Ev Idx": 9593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908412.900, "dur": 3.802,
    "args": {
      "Python parent id": 6654, "Python id": 6694, "finished": true, "Ev Idx": 9594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908417.300, "dur": 8.134,
    "args": {
      "Python parent id": 12, "Python id": 6695, "finished": true, "Ev Idx": 9595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908425.865, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 6696, "finished": true, "Ev Idx": 9596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908426.620, "dur": 0.234,
    "args": {
      "Python parent id": 12, "Python id": 6697, "finished": true, "Ev Idx": 9597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908427.371, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 6698, "finished": true, "Ev Idx": 9598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908427.835, "dur": 10.264,
    "args": {
      "Python parent id": 12, "Python id": 6699, "finished": true, "Ev Idx": 9599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908439.347, "dur": 14.208,
    "args": {
      "Python parent id": 12, "Python id": 6700, "finished": true, "Ev Idx": 9600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908439.671, "dur": 10.486,
    "args": {
      "Python parent id": 6700, "Python id": 6701, "finished": true, "Ev Idx": 9601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908439.817, "dur": 0.453,
    "args": {
      "Python parent id": 6701, "Python id": 6702, "finished": true, "Ev Idx": 9602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908439.963, "dur": 0.219,
    "args": {
      "Python parent id": 6702, "Python id": 6703, "finished": true, "Ev Idx": 9603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908440.056, "dur": 0.091,
    "args": {
      "Python parent id": 6703, "Python id": 6704, "finished": true, "Ev Idx": 9604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908440.583, "dur": 7.874,
    "args": {
      "Python parent id": 6701, "Python id": 6705, "finished": true, "Ev Idx": 9605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908440.667, "dur": 0.068,
    "args": {
      "Python parent id": 6705, "Python id": 6706, "finished": true, "Ev Idx": 9606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908440.843, "dur": 0.059,
    "args": {
      "Python parent id": 6705, "Python id": 6707, "finished": true, "Ev Idx": 9607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908441.069, "dur": 0.058,
    "args": {
      "Python parent id": 6705, "Python id": 6708, "finished": true, "Ev Idx": 9608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908441.549, "dur": 0.033,
    "args": {
      "Python parent id": 6705, "Python id": 6709, "finished": true, "Ev Idx": 9609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908441.851, "dur": 0.072,
    "args": {
      "Python parent id": 6705, "Python id": 6710, "finished": true, "Ev Idx": 9610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908442.241, "dur": 6.157,
    "args": {
      "Python parent id": 6705, "Python id": 6711, "finished": true, "Ev Idx": 9611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908442.325, "dur": 0.062,
    "args": {
      "Python parent id": 6711, "Python id": 6712, "finished": true, "Ev Idx": 9612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908442.524, "dur": 0.059,
    "args": {
      "Python parent id": 6711, "Python id": 6713, "finished": true, "Ev Idx": 9613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908442.751, "dur": 0.062,
    "args": {
      "Python parent id": 6711, "Python id": 6714, "finished": true, "Ev Idx": 9614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908443.081, "dur": 0.034,
    "args": {
      "Python parent id": 6711, "Python id": 6715, "finished": true, "Ev Idx": 9615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908443.414, "dur": 4.900,
    "args": {
      "Python parent id": 6711, "Python id": 6716, "finished": true, "Ev Idx": 9616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908443.635, "dur": 4.620,
    "args": {
      "Python parent id": 6716, "Python id": 6717, "finished": true, "Ev Idx": 9617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908443.755, "dur": 2.956,
    "args": {
      "Python parent id": 6717, "Python id": 6718, "finished": true, "Ev Idx": 9618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908443.954, "dur": 2.718,
    "args": {
      "Python parent id": 6718, "Python id": 6719, "finished": true, "Ev Idx": 9619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908444.072, "dur": 0.370,
    "args": {
      "Python parent id": 6719, "Python id": 6720, "finished": true, "Ev Idx": 9620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908444.312, "dur": 0.057,
    "args": {
      "Python parent id": 6720, "Python id": 6721, "finished": true, "Ev Idx": 9621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908444.618, "dur": 1.740,
    "args": {
      "Python parent id": 6719, "Python id": 6722, "finished": true, "Ev Idx": 9622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908444.783, "dur": 1.494,
    "args": {
      "Python parent id": 6722, "Python id": 6723, "finished": true, "Ev Idx": 9623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908444.954, "dur": 1.279,
    "args": {
      "Python parent id": 6723, "Python id": 6724, "finished": true, "Ev Idx": 9624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908445.095, "dur": 0.924,
    "args": {
      "Python parent id": 6724, "Python id": 6725, "finished": true, "Ev Idx": 9625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908445.241, "dur": 0.401,
    "args": {
      "Python parent id": 6725, "Python id": 6726, "finished": true, "Ev Idx": 9626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908445.320, "dur": 0.039,
    "args": {
      "Python parent id": 6726, "Python id": 6727, "finished": true, "Ev Idx": 9627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908445.488, "dur": 0.109,
    "args": {
      "Python parent id": 6726, "Python id": 6728, "finished": true, "Ev Idx": 9628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908446.579, "dur": 0.059,
    "args": {
      "Python parent id": 6719, "Python id": 6729, "finished": true, "Ev Idx": 9629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908446.849, "dur": 0.092,
    "args": {
      "Python parent id": 6717, "Python id": 6730, "finished": true, "Ev Idx": 9630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.169, "dur": 1.041,
    "args": {
      "Python parent id": 6717, "Python id": 6731, "finished": true, "Ev Idx": 9631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.303, "dur": 0.863,
    "args": {
      "Python parent id": 6731, "Python id": 6732, "finished": true, "Ev Idx": 9632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.432, "dur": 0.348,
    "args": {
      "Python parent id": 6732, "Python id": 6733, "finished": true, "Ev Idx": 9633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.559, "dur": 0.166,
    "args": {
      "Python parent id": 6733, "Python id": 6734, "finished": true, "Ev Idx": 9634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.637, "dur": 0.057,
    "args": {
      "Python parent id": 6734, "Python id": 6735, "finished": true, "Ev Idx": 9635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908447.902, "dur": 0.234,
    "args": {
      "Python parent id": 6732, "Python id": 6736, "finished": true, "Ev Idx": 9636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908448.569, "dur": 0.121,
    "args": {
      "Python parent id": 6701, "Python id": 6737, "finished": true, "Ev Idx": 9637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908449.224, "dur": 0.815,
    "args": {
      "Python parent id": 6701, "Python id": 6738, "finished": true, "Ev Idx": 9638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908449.583, "dur": 0.409,
    "args": {
      "Python parent id": 6738, "Python id": 6739, "finished": true, "Ev Idx": 9639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908450.430, "dur": 3.075,
    "args": {
      "Python parent id": 6700, "Python id": 6740, "finished": true, "Ev Idx": 9640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908454.399, "dur": 88.983,
    "args": {
      "Python parent id": 12, "Python id": 6741, "finished": true, "Ev Idx": 9641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908454.655, "dur": 0.398,
    "args": {
      "Python parent id": 6741, "Python id": 6742, "finished": true, "Ev Idx": 9642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908454.893, "dur": 0.080,
    "args": {
      "Python parent id": 6742, "Python id": 6743, "finished": true, "Ev Idx": 9643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908455.353, "dur": 0.328,
    "args": {
      "Python parent id": 6741, "Python id": 6744, "finished": true, "Ev Idx": 9644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908455.876, "dur": 0.267,
    "args": {
      "Python parent id": 6741, "Python id": 6745, "finished": true, "Ev Idx": 9645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908456.390, "dur": 1.503,
    "args": {
      "Python parent id": 6741, "Python id": 6746, "finished": true, "Ev Idx": 9646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908458.400, "dur": 84.844,
    "args": {
      "Python parent id": 6741, "Python id": 6747, "finished": true, "Ev Idx": 9647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908459.121, "dur": 83.924,
    "args": {
      "Python parent id": 6747, "Python id": 6748, "finished": true, "Ev Idx": 9648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908459.335, "dur": 83.634,
    "args": {
      "Python parent id": 6748, "Python id": 6749, "finished": true, "Ev Idx": 9649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908464.132, "dur": 77.623,
    "args": {
      "Python parent id": 6749, "Python id": 6750, "finished": true, "Ev Idx": 9650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908464.419, "dur": 77.262,
    "args": {
      "Python parent id": 6750, "Python id": 6751, "finished": true, "Ev Idx": 9651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908464.903, "dur": 76.625,
    "args": {
      "Python parent id": 6751, "Python id": 6752, "finished": true, "Ev Idx": 9652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908465.409, "dur": 0.339,
    "args": {
      "Python parent id": 6752, "Python id": 6753, "finished": true, "Ev Idx": 9653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908465.646, "dur": 0.047,
    "args": {
      "Python parent id": 6753, "Python id": 6754, "finished": true, "Ev Idx": 9654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908466.031, "dur": 0.299,
    "args": {
      "Python parent id": 6752, "Python id": 6755, "finished": true, "Ev Idx": 9655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908466.244, "dur": 0.035,
    "args": {
      "Python parent id": 6755, "Python id": 6756, "finished": true, "Ev Idx": 9656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908466.619, "dur": 74.844,
    "args": {
      "Python parent id": 6752, "Python id": 6757, "finished": true, "Ev Idx": 9657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908467.126, "dur": 73.119,
    "args": {
      "Python parent id": 6757, "Python id": 6758, "finished": true, "Ev Idx": 9658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908474.792, "dur": 18.869,
    "args": {
      "Python parent id": 6758, "Python id": 6759, "finished": true, "Ev Idx": 9659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908474.890, "dur": 0.116,
    "args": {
      "Python parent id": 6759, "Python id": 6760, "finished": true, "Ev Idx": 9660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908475.263, "dur": 18.272,
    "args": {
      "Python parent id": 6759, "Python id": 6761, "finished": true, "Ev Idx": 9661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908475.385, "dur": 8.485,
    "args": {
      "Python parent id": 6761, "Python id": 6762, "finished": true, "Ev Idx": 9662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908501.840, "dur": 6.982,
    "args": {
      "Python parent id": 6758, "Python id": 6763, "finished": true, "Ev Idx": 9663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908516.617, "dur": 8.132,
    "args": {
      "Python parent id": 6758, "Python id": 6764, "finished": true, "Ev Idx": 9664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908544.609, "dur": 15.378,
    "args": {
      "Python parent id": 12, "Python id": 6765, "finished": true, "Ev Idx": 9665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908544.932, "dur": 11.298,
    "args": {
      "Python parent id": 6765, "Python id": 6766, "finished": true, "Ev Idx": 9666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908545.164, "dur": 0.425,
    "args": {
      "Python parent id": 6766, "Python id": 6767, "finished": true, "Ev Idx": 9667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908545.315, "dur": 0.191,
    "args": {
      "Python parent id": 6767, "Python id": 6768, "finished": true, "Ev Idx": 9668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908545.396, "dur": 0.074,
    "args": {
      "Python parent id": 6768, "Python id": 6769, "finished": true, "Ev Idx": 9669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908545.843, "dur": 8.382,
    "args": {
      "Python parent id": 6766, "Python id": 6770, "finished": true, "Ev Idx": 9670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908545.919, "dur": 0.100,
    "args": {
      "Python parent id": 6770, "Python id": 6771, "finished": true, "Ev Idx": 9671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908546.113, "dur": 0.056,
    "args": {
      "Python parent id": 6770, "Python id": 6772, "finished": true, "Ev Idx": 9672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908546.300, "dur": 0.059,
    "args": {
      "Python parent id": 6770, "Python id": 6773, "finished": true, "Ev Idx": 9673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908546.781, "dur": 0.042,
    "args": {
      "Python parent id": 6770, "Python id": 6774, "finished": true, "Ev Idx": 9674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908547.122, "dur": 0.067,
    "args": {
      "Python parent id": 6770, "Python id": 6775, "finished": true, "Ev Idx": 9675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908547.550, "dur": 6.594,
    "args": {
      "Python parent id": 6770, "Python id": 6776, "finished": true, "Ev Idx": 9676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908547.645, "dur": 0.058,
    "args": {
      "Python parent id": 6776, "Python id": 6777, "finished": true, "Ev Idx": 9677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908547.857, "dur": 0.057,
    "args": {
      "Python parent id": 6776, "Python id": 6778, "finished": true, "Ev Idx": 9678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908548.146, "dur": 0.066,
    "args": {
      "Python parent id": 6776, "Python id": 6779, "finished": true, "Ev Idx": 9679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908548.554, "dur": 0.060,
    "args": {
      "Python parent id": 6776, "Python id": 6780, "finished": true, "Ev Idx": 9680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908548.932, "dur": 5.137,
    "args": {
      "Python parent id": 6776, "Python id": 6781, "finished": true, "Ev Idx": 9681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908549.206, "dur": 4.789,
    "args": {
      "Python parent id": 6781, "Python id": 6782, "finished": true, "Ev Idx": 9682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908549.323, "dur": 3.067,
    "args": {
      "Python parent id": 6782, "Python id": 6783, "finished": true, "Ev Idx": 9683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908549.513, "dur": 2.838,
    "args": {
      "Python parent id": 6783, "Python id": 6784, "finished": true, "Ev Idx": 9684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908549.633, "dur": 0.276,
    "args": {
      "Python parent id": 6784, "Python id": 6785, "finished": true, "Ev Idx": 9685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908549.781, "dur": 0.056,
    "args": {
      "Python parent id": 6785, "Python id": 6786, "finished": true, "Ev Idx": 9686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.065, "dur": 1.974,
    "args": {
      "Python parent id": 6784, "Python id": 6787, "finished": true, "Ev Idx": 9687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.273, "dur": 1.691,
    "args": {
      "Python parent id": 6787, "Python id": 6788, "finished": true, "Ev Idx": 9688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.474, "dur": 1.448,
    "args": {
      "Python parent id": 6788, "Python id": 6789, "finished": true, "Ev Idx": 9689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.609, "dur": 1.104,
    "args": {
      "Python parent id": 6789, "Python id": 6790, "finished": true, "Ev Idx": 9690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.770, "dur": 0.441,
    "args": {
      "Python parent id": 6790, "Python id": 6791, "finished": true, "Ev Idx": 9691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908550.849, "dur": 0.040,
    "args": {
      "Python parent id": 6791, "Python id": 6792, "finished": true, "Ev Idx": 9692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908551.041, "dur": 0.132,
    "args": {
      "Python parent id": 6791, "Python id": 6793, "finished": true, "Ev Idx": 9693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908552.196, "dur": 0.117,
    "args": {
      "Python parent id": 6784, "Python id": 6794, "finished": true, "Ev Idx": 9694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908552.565, "dur": 0.074,
    "args": {
      "Python parent id": 6782, "Python id": 6795, "finished": true, "Ev Idx": 9695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908552.864, "dur": 1.091,
    "args": {
      "Python parent id": 6782, "Python id": 6796, "finished": true, "Ev Idx": 9696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908553.039, "dur": 0.873,
    "args": {
      "Python parent id": 6796, "Python id": 6797, "finished": true, "Ev Idx": 9697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908553.159, "dur": 0.377,
    "args": {
      "Python parent id": 6797, "Python id": 6798, "finished": true, "Ev Idx": 9698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908553.311, "dur": 0.166,
    "args": {
      "Python parent id": 6798, "Python id": 6799, "finished": true, "Ev Idx": 9699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908553.394, "dur": 0.055,
    "args": {
      "Python parent id": 6799, "Python id": 6800, "finished": true, "Ev Idx": 9700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908553.662, "dur": 0.224,
    "args": {
      "Python parent id": 6797, "Python id": 6801, "finished": true, "Ev Idx": 9701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908554.362, "dur": 0.205,
    "args": {
      "Python parent id": 6766, "Python id": 6802, "finished": true, "Ev Idx": 9702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908555.091, "dur": 1.014,
    "args": {
      "Python parent id": 6766, "Python id": 6803, "finished": true, "Ev Idx": 9703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908555.606, "dur": 0.452,
    "args": {
      "Python parent id": 6803, "Python id": 6804, "finished": true, "Ev Idx": 9704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908556.522, "dur": 3.393,
    "args": {
      "Python parent id": 6765, "Python id": 6805, "finished": true, "Ev Idx": 9705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908560.481, "dur": 7.710,
    "args": {
      "Python parent id": 12, "Python id": 6806, "finished": true, "Ev Idx": 9706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908568.662, "dur": 0.115,
    "args": {
      "Python parent id": 12, "Python id": 6807, "finished": true, "Ev Idx": 9707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908569.437, "dur": 0.199,
    "args": {
      "Python parent id": 12, "Python id": 6808, "finished": true, "Ev Idx": 9708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908570.090, "dur": 0.070,
    "args": {
      "Python parent id": 12, "Python id": 6809, "finished": true, "Ev Idx": 9709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908570.499, "dur": 9.835,
    "args": {
      "Python parent id": 12, "Python id": 6810, "finished": true, "Ev Idx": 9710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908581.603, "dur": 13.949,
    "args": {
      "Python parent id": 12, "Python id": 6811, "finished": true, "Ev Idx": 9711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908581.865, "dur": 10.114,
    "args": {
      "Python parent id": 6811, "Python id": 6812, "finished": true, "Ev Idx": 9712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.002, "dur": 0.447,
    "args": {
      "Python parent id": 6812, "Python id": 6813, "finished": true, "Ev Idx": 9713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.147, "dur": 0.228,
    "args": {
      "Python parent id": 6813, "Python id": 6814, "finished": true, "Ev Idx": 9714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.241, "dur": 0.101,
    "args": {
      "Python parent id": 6814, "Python id": 6815, "finished": true, "Ev Idx": 9715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.713, "dur": 7.625,
    "args": {
      "Python parent id": 6812, "Python id": 6816, "finished": true, "Ev Idx": 9716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.792, "dur": 0.062,
    "args": {
      "Python parent id": 6816, "Python id": 6817, "finished": true, "Ev Idx": 9717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908582.965, "dur": 0.056,
    "args": {
      "Python parent id": 6816, "Python id": 6818, "finished": true, "Ev Idx": 9718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908583.160, "dur": 0.057,
    "args": {
      "Python parent id": 6816, "Python id": 6819, "finished": true, "Ev Idx": 9719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908583.599, "dur": 0.059,
    "args": {
      "Python parent id": 6816, "Python id": 6820, "finished": true, "Ev Idx": 9720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908583.892, "dur": 0.070,
    "args": {
      "Python parent id": 6816, "Python id": 6821, "finished": true, "Ev Idx": 9721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908584.261, "dur": 5.981,
    "args": {
      "Python parent id": 6816, "Python id": 6822, "finished": true, "Ev Idx": 9722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908584.335, "dur": 0.054,
    "args": {
      "Python parent id": 6822, "Python id": 6823, "finished": true, "Ev Idx": 9723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908584.525, "dur": 0.051,
    "args": {
      "Python parent id": 6822, "Python id": 6824, "finished": true, "Ev Idx": 9724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908584.761, "dur": 0.064,
    "args": {
      "Python parent id": 6822, "Python id": 6825, "finished": true, "Ev Idx": 9725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908585.142, "dur": 0.033,
    "args": {
      "Python parent id": 6822, "Python id": 6826, "finished": true, "Ev Idx": 9726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908585.479, "dur": 4.708,
    "args": {
      "Python parent id": 6822, "Python id": 6827, "finished": true, "Ev Idx": 9727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908585.671, "dur": 4.461,
    "args": {
      "Python parent id": 6827, "Python id": 6828, "finished": true, "Ev Idx": 9728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908585.797, "dur": 2.785,
    "args": {
      "Python parent id": 6828, "Python id": 6829, "finished": true, "Ev Idx": 9729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908585.957, "dur": 2.580,
    "args": {
      "Python parent id": 6829, "Python id": 6830, "finished": true, "Ev Idx": 9730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.075, "dur": 0.248,
    "args": {
      "Python parent id": 6830, "Python id": 6831, "finished": true, "Ev Idx": 9731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.205, "dur": 0.058,
    "args": {
      "Python parent id": 6831, "Python id": 6832, "finished": true, "Ev Idx": 9732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.496, "dur": 1.745,
    "args": {
      "Python parent id": 6830, "Python id": 6833, "finished": true, "Ev Idx": 9733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.680, "dur": 1.494,
    "args": {
      "Python parent id": 6833, "Python id": 6834, "finished": true, "Ev Idx": 9734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.884, "dur": 1.243,
    "args": {
      "Python parent id": 6834, "Python id": 6835, "finished": true, "Ev Idx": 9735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908586.995, "dur": 0.936,
    "args": {
      "Python parent id": 6835, "Python id": 6836, "finished": true, "Ev Idx": 9736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908587.140, "dur": 0.408,
    "args": {
      "Python parent id": 6836, "Python id": 6837, "finished": true, "Ev Idx": 9737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908587.215, "dur": 0.042,
    "args": {
      "Python parent id": 6837, "Python id": 6838, "finished": true, "Ev Idx": 9738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908587.389, "dur": 0.123,
    "args": {
      "Python parent id": 6837, "Python id": 6839, "finished": true, "Ev Idx": 9739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908588.405, "dur": 0.102,
    "args": {
      "Python parent id": 6830, "Python id": 6840, "finished": true, "Ev Idx": 9740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908588.735, "dur": 0.074,
    "args": {
      "Python parent id": 6828, "Python id": 6841, "finished": true, "Ev Idx": 9741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.029, "dur": 1.062,
    "args": {
      "Python parent id": 6828, "Python id": 6842, "finished": true, "Ev Idx": 9742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.150, "dur": 0.899,
    "args": {
      "Python parent id": 6842, "Python id": 6843, "finished": true, "Ev Idx": 9743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.301, "dur": 0.386,
    "args": {
      "Python parent id": 6843, "Python id": 6844, "finished": true, "Ev Idx": 9744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.464, "dur": 0.165,
    "args": {
      "Python parent id": 6844, "Python id": 6845, "finished": true, "Ev Idx": 9745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.541, "dur": 0.058,
    "args": {
      "Python parent id": 6845, "Python id": 6846, "finished": true, "Ev Idx": 9746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908589.807, "dur": 0.214,
    "args": {
      "Python parent id": 6843, "Python id": 6847, "finished": true, "Ev Idx": 9747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908590.444, "dur": 0.143,
    "args": {
      "Python parent id": 6812, "Python id": 6848, "finished": true, "Ev Idx": 9748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908591.059, "dur": 0.805,
    "args": {
      "Python parent id": 6812, "Python id": 6849, "finished": true, "Ev Idx": 9749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908591.407, "dur": 0.410,
    "args": {
      "Python parent id": 6849, "Python id": 6850, "finished": true, "Ev Idx": 9750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908592.202, "dur": 3.276,
    "args": {
      "Python parent id": 6811, "Python id": 6851, "finished": true, "Ev Idx": 9751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908596.485, "dur": 85.808,
    "args": {
      "Python parent id": 12, "Python id": 6852, "finished": true, "Ev Idx": 9752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908596.741, "dur": 0.421,
    "args": {
      "Python parent id": 6852, "Python id": 6853, "finished": true, "Ev Idx": 9753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908596.983, "dur": 0.075,
    "args": {
      "Python parent id": 6853, "Python id": 6854, "finished": true, "Ev Idx": 9754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908597.458, "dur": 0.352,
    "args": {
      "Python parent id": 6852, "Python id": 6855, "finished": true, "Ev Idx": 9755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908598.005, "dur": 0.292,
    "args": {
      "Python parent id": 6852, "Python id": 6856, "finished": true, "Ev Idx": 9756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908598.516, "dur": 1.322,
    "args": {
      "Python parent id": 6852, "Python id": 6857, "finished": true, "Ev Idx": 9757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908600.342, "dur": 81.833,
    "args": {
      "Python parent id": 6852, "Python id": 6858, "finished": true, "Ev Idx": 9758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908601.014, "dur": 81.010,
    "args": {
      "Python parent id": 6858, "Python id": 6859, "finished": true, "Ev Idx": 9759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908601.252, "dur": 80.694,
    "args": {
      "Python parent id": 6859, "Python id": 6860, "finished": true, "Ev Idx": 9760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908606.053, "dur": 74.727,
    "args": {
      "Python parent id": 6860, "Python id": 6861, "finished": true, "Ev Idx": 9761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908606.334, "dur": 74.354,
    "args": {
      "Python parent id": 6861, "Python id": 6862, "finished": true, "Ev Idx": 9762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908606.891, "dur": 73.711,
    "args": {
      "Python parent id": 6862, "Python id": 6863, "finished": true, "Ev Idx": 9763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908607.434, "dur": 0.372,
    "args": {
      "Python parent id": 6863, "Python id": 6864, "finished": true, "Ev Idx": 9764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908607.698, "dur": 0.040,
    "args": {
      "Python parent id": 6864, "Python id": 6865, "finished": true, "Ev Idx": 9765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908608.123, "dur": 0.283,
    "args": {
      "Python parent id": 6863, "Python id": 6866, "finished": true, "Ev Idx": 9766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908608.321, "dur": 0.032,
    "args": {
      "Python parent id": 6866, "Python id": 6867, "finished": true, "Ev Idx": 9767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908608.699, "dur": 71.813,
    "args": {
      "Python parent id": 6863, "Python id": 6868, "finished": true, "Ev Idx": 9768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908609.362, "dur": 69.981,
    "args": {
      "Python parent id": 6868, "Python id": 6869, "finished": true, "Ev Idx": 9769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908615.930, "dur": 17.727,
    "args": {
      "Python parent id": 6869, "Python id": 6870, "finished": true, "Ev Idx": 9770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908616.053, "dur": 0.083,
    "args": {
      "Python parent id": 6870, "Python id": 6871, "finished": true, "Ev Idx": 9771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908616.359, "dur": 17.195,
    "args": {
      "Python parent id": 6870, "Python id": 6872, "finished": true, "Ev Idx": 9772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908616.496, "dur": 8.074,
    "args": {
      "Python parent id": 6872, "Python id": 6873, "finished": true, "Ev Idx": 9773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908642.106, "dur": 6.814,
    "args": {
      "Python parent id": 6869, "Python id": 6874, "finished": true, "Ev Idx": 9774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908656.836, "dur": 7.954,
    "args": {
      "Python parent id": 6869, "Python id": 6875, "finished": true, "Ev Idx": 9775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908683.474, "dur": 15.508,
    "args": {
      "Python parent id": 12, "Python id": 6876, "finished": true, "Ev Idx": 9776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908683.866, "dur": 11.532,
    "args": {
      "Python parent id": 6876, "Python id": 6877, "finished": true, "Ev Idx": 9777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908684.078, "dur": 0.412,
    "args": {
      "Python parent id": 6877, "Python id": 6878, "finished": true, "Ev Idx": 9778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908684.230, "dur": 0.202,
    "args": {
      "Python parent id": 6878, "Python id": 6879, "finished": true, "Ev Idx": 9779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908684.322, "dur": 0.074,
    "args": {
      "Python parent id": 6879, "Python id": 6880, "finished": true, "Ev Idx": 9780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908684.780, "dur": 8.511,
    "args": {
      "Python parent id": 6877, "Python id": 6881, "finished": true, "Ev Idx": 9781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908684.856, "dur": 0.084,
    "args": {
      "Python parent id": 6881, "Python id": 6882, "finished": true, "Ev Idx": 9782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908685.054, "dur": 0.054,
    "args": {
      "Python parent id": 6881, "Python id": 6883, "finished": true, "Ev Idx": 9783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908685.251, "dur": 0.054,
    "args": {
      "Python parent id": 6881, "Python id": 6884, "finished": true, "Ev Idx": 9784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908685.786, "dur": 0.053,
    "args": {
      "Python parent id": 6881, "Python id": 6885, "finished": true, "Ev Idx": 9785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908686.089, "dur": 0.069,
    "args": {
      "Python parent id": 6881, "Python id": 6886, "finished": true, "Ev Idx": 9786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908686.489, "dur": 6.724,
    "args": {
      "Python parent id": 6881, "Python id": 6887, "finished": true, "Ev Idx": 9787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908686.590, "dur": 0.056,
    "args": {
      "Python parent id": 6887, "Python id": 6888, "finished": true, "Ev Idx": 9788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908686.797, "dur": 0.049,
    "args": {
      "Python parent id": 6887, "Python id": 6889, "finished": true, "Ev Idx": 9789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908687.016, "dur": 0.058,
    "args": {
      "Python parent id": 6887, "Python id": 6890, "finished": true, "Ev Idx": 9790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908687.417, "dur": 0.049,
    "args": {
      "Python parent id": 6887, "Python id": 6891, "finished": true, "Ev Idx": 9791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908687.787, "dur": 5.310,
    "args": {
      "Python parent id": 6887, "Python id": 6892, "finished": true, "Ev Idx": 9792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908687.989, "dur": 5.022,
    "args": {
      "Python parent id": 6892, "Python id": 6893, "finished": true, "Ev Idx": 9793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908688.116, "dur": 3.199,
    "args": {
      "Python parent id": 6893, "Python id": 6894, "finished": true, "Ev Idx": 9794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908688.328, "dur": 2.919,
    "args": {
      "Python parent id": 6894, "Python id": 6895, "finished": true, "Ev Idx": 9795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908688.484, "dur": 0.271,
    "args": {
      "Python parent id": 6895, "Python id": 6896, "finished": true, "Ev Idx": 9796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908688.599, "dur": 0.084,
    "args": {
      "Python parent id": 6896, "Python id": 6897, "finished": true, "Ev Idx": 9797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908688.885, "dur": 1.978,
    "args": {
      "Python parent id": 6895, "Python id": 6898, "finished": true, "Ev Idx": 9798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.102, "dur": 1.683,
    "args": {
      "Python parent id": 6898, "Python id": 6899, "finished": true, "Ev Idx": 9799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.285, "dur": 1.439,
    "args": {
      "Python parent id": 6899, "Python id": 6900, "finished": true, "Ev Idx": 9800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.403, "dur": 1.141,
    "args": {
      "Python parent id": 6900, "Python id": 6901, "finished": true, "Ev Idx": 9801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.576, "dur": 0.497,
    "args": {
      "Python parent id": 6901, "Python id": 6902, "finished": true, "Ev Idx": 9802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.684, "dur": 0.046,
    "args": {
      "Python parent id": 6902, "Python id": 6903, "finished": true, "Ev Idx": 9803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908689.887, "dur": 0.125,
    "args": {
      "Python parent id": 6902, "Python id": 6904, "finished": true, "Ev Idx": 9804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908691.032, "dur": 0.167,
    "args": {
      "Python parent id": 6895, "Python id": 6905, "finished": true, "Ev Idx": 9805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908691.475, "dur": 0.090,
    "args": {
      "Python parent id": 6893, "Python id": 6906, "finished": true, "Ev Idx": 9806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908691.801, "dur": 1.171,
    "args": {
      "Python parent id": 6893, "Python id": 6907, "finished": true, "Ev Idx": 9807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908691.987, "dur": 0.947,
    "args": {
      "Python parent id": 6907, "Python id": 6908, "finished": true, "Ev Idx": 9808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908692.152, "dur": 0.349,
    "args": {
      "Python parent id": 6908, "Python id": 6909, "finished": true, "Ev Idx": 9809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908692.268, "dur": 0.164,
    "args": {
      "Python parent id": 6909, "Python id": 6910, "finished": true, "Ev Idx": 9810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908692.352, "dur": 0.050,
    "args": {
      "Python parent id": 6910, "Python id": 6911, "finished": true, "Ev Idx": 9811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908692.619, "dur": 0.269,
    "args": {
      "Python parent id": 6908, "Python id": 6912, "finished": true, "Ev Idx": 9812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908693.428, "dur": 0.179,
    "args": {
      "Python parent id": 6877, "Python id": 6913, "finished": true, "Ev Idx": 9813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908694.180, "dur": 1.094,
    "args": {
      "Python parent id": 6877, "Python id": 6914, "finished": true, "Ev Idx": 9814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908694.636, "dur": 0.590,
    "args": {
      "Python parent id": 6914, "Python id": 6915, "finished": true, "Ev Idx": 9815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908695.711, "dur": 3.216,
    "args": {
      "Python parent id": 6876, "Python id": 6916, "finished": true, "Ev Idx": 9816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908699.456, "dur": 7.827,
    "args": {
      "Python parent id": 12, "Python id": 6917, "finished": true, "Ev Idx": 9817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908707.693, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 6918, "finished": true, "Ev Idx": 9818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908708.363, "dur": 0.213,
    "args": {
      "Python parent id": 12, "Python id": 6919, "finished": true, "Ev Idx": 9819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908709.012, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 6920, "finished": true, "Ev Idx": 9820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908709.449, "dur": 9.397,
    "args": {
      "Python parent id": 12, "Python id": 6921, "finished": true, "Ev Idx": 9821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908720.032, "dur": 14.059,
    "args": {
      "Python parent id": 12, "Python id": 6922, "finished": true, "Ev Idx": 9822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908720.295, "dur": 10.650,
    "args": {
      "Python parent id": 6922, "Python id": 6923, "finished": true, "Ev Idx": 9823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908720.447, "dur": 0.491,
    "args": {
      "Python parent id": 6923, "Python id": 6924, "finished": true, "Ev Idx": 9824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908720.639, "dur": 0.216,
    "args": {
      "Python parent id": 6924, "Python id": 6925, "finished": true, "Ev Idx": 9825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908720.734, "dur": 0.089,
    "args": {
      "Python parent id": 6925, "Python id": 6926, "finished": true, "Ev Idx": 9826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908721.214, "dur": 8.004,
    "args": {
      "Python parent id": 6923, "Python id": 6927, "finished": true, "Ev Idx": 9827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908721.329, "dur": 0.068,
    "args": {
      "Python parent id": 6927, "Python id": 6928, "finished": true, "Ev Idx": 9828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908721.512, "dur": 0.060,
    "args": {
      "Python parent id": 6927, "Python id": 6929, "finished": true, "Ev Idx": 9829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908721.713, "dur": 0.062,
    "args": {
      "Python parent id": 6927, "Python id": 6930, "finished": true, "Ev Idx": 9830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908722.178, "dur": 0.036,
    "args": {
      "Python parent id": 6927, "Python id": 6931, "finished": true, "Ev Idx": 9831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908722.450, "dur": 0.072,
    "args": {
      "Python parent id": 6927, "Python id": 6932, "finished": true, "Ev Idx": 9832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908722.868, "dur": 6.297,
    "args": {
      "Python parent id": 6927, "Python id": 6933, "finished": true, "Ev Idx": 9833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908723.008, "dur": 0.059,
    "args": {
      "Python parent id": 6933, "Python id": 6934, "finished": true, "Ev Idx": 9834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908723.211, "dur": 0.063,
    "args": {
      "Python parent id": 6933, "Python id": 6935, "finished": true, "Ev Idx": 9835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908723.467, "dur": 0.067,
    "args": {
      "Python parent id": 6933, "Python id": 6936, "finished": true, "Ev Idx": 9836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908723.849, "dur": 0.044,
    "args": {
      "Python parent id": 6933, "Python id": 6937, "finished": true, "Ev Idx": 9837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908724.212, "dur": 4.846,
    "args": {
      "Python parent id": 6933, "Python id": 6938, "finished": true, "Ev Idx": 9838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908724.484, "dur": 4.519,
    "args": {
      "Python parent id": 6938, "Python id": 6939, "finished": true, "Ev Idx": 9839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908724.656, "dur": 2.801,
    "args": {
      "Python parent id": 6939, "Python id": 6940, "finished": true, "Ev Idx": 9840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908724.829, "dur": 2.584,
    "args": {
      "Python parent id": 6940, "Python id": 6941, "finished": true, "Ev Idx": 9841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908724.970, "dur": 0.274,
    "args": {
      "Python parent id": 6941, "Python id": 6942, "finished": true, "Ev Idx": 9842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908725.108, "dur": 0.063,
    "args": {
      "Python parent id": 6942, "Python id": 6943, "finished": true, "Ev Idx": 9843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908725.422, "dur": 1.729,
    "args": {
      "Python parent id": 6941, "Python id": 6944, "finished": true, "Ev Idx": 9844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908725.587, "dur": 1.498,
    "args": {
      "Python parent id": 6944, "Python id": 6945, "finished": true, "Ev Idx": 9845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908725.751, "dur": 1.292,
    "args": {
      "Python parent id": 6945, "Python id": 6946, "finished": true, "Ev Idx": 9846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908725.889, "dur": 0.991,
    "args": {
      "Python parent id": 6946, "Python id": 6947, "finished": true, "Ev Idx": 9847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908726.034, "dur": 0.451,
    "args": {
      "Python parent id": 6947, "Python id": 6948, "finished": true, "Ev Idx": 9848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908726.109, "dur": 0.048,
    "args": {
      "Python parent id": 6948, "Python id": 6949, "finished": true, "Ev Idx": 9849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908726.306, "dur": 0.137,
    "args": {
      "Python parent id": 6948, "Python id": 6950, "finished": true, "Ev Idx": 9850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908727.308, "dur": 0.074,
    "args": {
      "Python parent id": 6941, "Python id": 6951, "finished": true, "Ev Idx": 9851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908727.602, "dur": 0.075,
    "args": {
      "Python parent id": 6939, "Python id": 6952, "finished": true, "Ev Idx": 9852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908727.931, "dur": 1.033,
    "args": {
      "Python parent id": 6939, "Python id": 6953, "finished": true, "Ev Idx": 9853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908728.055, "dur": 0.866,
    "args": {
      "Python parent id": 6953, "Python id": 6954, "finished": true, "Ev Idx": 9854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908728.187, "dur": 0.350,
    "args": {
      "Python parent id": 6954, "Python id": 6955, "finished": true, "Ev Idx": 9855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908728.315, "dur": 0.170,
    "args": {
      "Python parent id": 6955, "Python id": 6956, "finished": true, "Ev Idx": 9856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908728.400, "dur": 0.056,
    "args": {
      "Python parent id": 6956, "Python id": 6957, "finished": true, "Ev Idx": 9857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908728.658, "dur": 0.230,
    "args": {
      "Python parent id": 6954, "Python id": 6958, "finished": true, "Ev Idx": 9858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908729.334, "dur": 0.159,
    "args": {
      "Python parent id": 6923, "Python id": 6959, "finished": true, "Ev Idx": 9859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908730.003, "dur": 0.828,
    "args": {
      "Python parent id": 6923, "Python id": 6960, "finished": true, "Ev Idx": 9860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908730.365, "dur": 0.417,
    "args": {
      "Python parent id": 6960, "Python id": 6961, "finished": true, "Ev Idx": 9861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908731.208, "dur": 2.825,
    "args": {
      "Python parent id": 6922, "Python id": 6962, "finished": true, "Ev Idx": 9862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908734.923, "dur": 84.784,
    "args": {
      "Python parent id": 12, "Python id": 6963, "finished": true, "Ev Idx": 9863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908735.171, "dur": 0.409,
    "args": {
      "Python parent id": 6963, "Python id": 6964, "finished": true, "Ev Idx": 9864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908735.419, "dur": 0.089,
    "args": {
      "Python parent id": 6964, "Python id": 6965, "finished": true, "Ev Idx": 9865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908735.811, "dur": 0.339,
    "args": {
      "Python parent id": 6963, "Python id": 6966, "finished": true, "Ev Idx": 9866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908736.373, "dur": 0.250,
    "args": {
      "Python parent id": 6963, "Python id": 6967, "finished": true, "Ev Idx": 9867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908736.819, "dur": 1.333,
    "args": {
      "Python parent id": 6963, "Python id": 6968, "finished": true, "Ev Idx": 9868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908738.706, "dur": 80.869,
    "args": {
      "Python parent id": 6963, "Python id": 6969, "finished": true, "Ev Idx": 9869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908739.473, "dur": 79.920,
    "args": {
      "Python parent id": 6969, "Python id": 6970, "finished": true, "Ev Idx": 9870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908739.694, "dur": 79.586,
    "args": {
      "Python parent id": 6970, "Python id": 6971, "finished": true, "Ev Idx": 9871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908744.450, "dur": 73.657,
    "args": {
      "Python parent id": 6971, "Python id": 6972, "finished": true, "Ev Idx": 9872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908744.735, "dur": 73.280,
    "args": {
      "Python parent id": 6972, "Python id": 6973, "finished": true, "Ev Idx": 9873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908745.231, "dur": 72.713,
    "args": {
      "Python parent id": 6973, "Python id": 6974, "finished": true, "Ev Idx": 9874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908745.732, "dur": 0.390,
    "args": {
      "Python parent id": 6974, "Python id": 6975, "finished": true, "Ev Idx": 9875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908746.005, "dur": 0.062,
    "args": {
      "Python parent id": 6975, "Python id": 6976, "finished": true, "Ev Idx": 9876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908746.419, "dur": 0.367,
    "args": {
      "Python parent id": 6974, "Python id": 6977, "finished": true, "Ev Idx": 9877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908746.679, "dur": 0.066,
    "args": {
      "Python parent id": 6977, "Python id": 6978, "finished": true, "Ev Idx": 9878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908747.121, "dur": 70.750,
    "args": {
      "Python parent id": 6974, "Python id": 6979, "finished": true, "Ev Idx": 9879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908747.749, "dur": 68.972,
    "args": {
      "Python parent id": 6979, "Python id": 6980, "finished": true, "Ev Idx": 9880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908754.310, "dur": 17.528,
    "args": {
      "Python parent id": 6980, "Python id": 6981, "finished": true, "Ev Idx": 9881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908754.432, "dur": 0.169,
    "args": {
      "Python parent id": 6981, "Python id": 6982, "finished": true, "Ev Idx": 9882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908754.835, "dur": 16.887,
    "args": {
      "Python parent id": 6981, "Python id": 6983, "finished": true, "Ev Idx": 9883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908754.963, "dur": 7.668,
    "args": {
      "Python parent id": 6983, "Python id": 6984, "finished": true, "Ev Idx": 9884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908779.328, "dur": 6.790,
    "args": {
      "Python parent id": 6980, "Python id": 6985, "finished": true, "Ev Idx": 9885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908794.819, "dur": 7.784,
    "args": {
      "Python parent id": 6980, "Python id": 6986, "finished": true, "Ev Idx": 9886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908820.764, "dur": 29.230,
    "args": {
      "Python parent id": 12, "Python id": 6987, "finished": true, "Ev Idx": 9887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908821.109, "dur": 24.253,
    "args": {
      "Python parent id": 6987, "Python id": 6988, "finished": true, "Ev Idx": 9888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908821.300, "dur": 0.555,
    "args": {
      "Python parent id": 6988, "Python id": 6989, "finished": true, "Ev Idx": 9889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908821.517, "dur": 0.247,
    "args": {
      "Python parent id": 6989, "Python id": 6990, "finished": true, "Ev Idx": 9890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908821.639, "dur": 0.067,
    "args": {
      "Python parent id": 6990, "Python id": 6991, "finished": true, "Ev Idx": 9891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908822.129, "dur": 20.223,
    "args": {
      "Python parent id": 6988, "Python id": 6992, "finished": true, "Ev Idx": 9892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908822.266, "dur": 0.078,
    "args": {
      "Python parent id": 6992, "Python id": 6993, "finished": true, "Ev Idx": 9893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908822.453, "dur": 0.058,
    "args": {
      "Python parent id": 6992, "Python id": 6994, "finished": true, "Ev Idx": 9894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908822.663, "dur": 0.059,
    "args": {
      "Python parent id": 6992, "Python id": 6995, "finished": true, "Ev Idx": 9895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908823.172, "dur": 0.028,
    "args": {
      "Python parent id": 6992, "Python id": 6996, "finished": true, "Ev Idx": 9896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908823.484, "dur": 0.071,
    "args": {
      "Python parent id": 6992, "Python id": 6997, "finished": true, "Ev Idx": 9897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908823.978, "dur": 18.302,
    "args": {
      "Python parent id": 6992, "Python id": 6998, "finished": true, "Ev Idx": 9898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908824.087, "dur": 0.065,
    "args": {
      "Python parent id": 6998, "Python id": 6999, "finished": true, "Ev Idx": 9899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908824.339, "dur": 0.053,
    "args": {
      "Python parent id": 6998, "Python id": 7000, "finished": true, "Ev Idx": 9900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908824.578, "dur": 0.063,
    "args": {
      "Python parent id": 6998, "Python id": 7001, "finished": true, "Ev Idx": 9901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908825.004, "dur": 0.063,
    "args": {
      "Python parent id": 6998, "Python id": 7002, "finished": true, "Ev Idx": 9902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908825.380, "dur": 16.856,
    "args": {
      "Python parent id": 6998, "Python id": 7003, "finished": true, "Ev Idx": 9903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908825.669, "dur": 16.460,
    "args": {
      "Python parent id": 7003, "Python id": 7004, "finished": true, "Ev Idx": 9904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908825.796, "dur": 3.374,
    "args": {
      "Python parent id": 7004, "Python id": 7005, "finished": true, "Ev Idx": 9905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908826.021, "dur": 3.080,
    "args": {
      "Python parent id": 7005, "Python id": 7006, "finished": true, "Ev Idx": 9906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908826.155, "dur": 0.260,
    "args": {
      "Python parent id": 7006, "Python id": 7007, "finished": true, "Ev Idx": 9907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908826.287, "dur": 0.061,
    "args": {
      "Python parent id": 7007, "Python id": 7008, "finished": true, "Ev Idx": 9908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908826.581, "dur": 2.149,
    "args": {
      "Python parent id": 7006, "Python id": 7009, "finished": true, "Ev Idx": 9909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908826.794, "dur": 1.861,
    "args": {
      "Python parent id": 7009, "Python id": 7010, "finished": true, "Ev Idx": 9910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908827.003, "dur": 1.608,
    "args": {
      "Python parent id": 7010, "Python id": 7011, "finished": true, "Ev Idx": 9911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908827.173, "dur": 1.223,
    "args": {
      "Python parent id": 7011, "Python id": 7012, "finished": true, "Ev Idx": 9912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908827.451, "dur": 0.499,
    "args": {
      "Python parent id": 7012, "Python id": 7013, "finished": true, "Ev Idx": 9913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908827.533, "dur": 0.044,
    "args": {
      "Python parent id": 7013, "Python id": 7014, "finished": true, "Ev Idx": 9914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908827.739, "dur": 0.132,
    "args": {
      "Python parent id": 7013, "Python id": 7015, "finished": true, "Ev Idx": 9915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908828.886, "dur": 0.184,
    "args": {
      "Python parent id": 7006, "Python id": 7016, "finished": true, "Ev Idx": 9916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908839.678, "dur": 0.205,
    "args": {
      "Python parent id": 7004, "Python id": 7017, "finished": true, "Ev Idx": 9917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908840.397, "dur": 1.671,
    "args": {
      "Python parent id": 7004, "Python id": 7018, "finished": true, "Ev Idx": 9918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908840.591, "dur": 1.434,
    "args": {
      "Python parent id": 7018, "Python id": 7019, "finished": true, "Ev Idx": 9919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908840.776, "dur": 0.572,
    "args": {
      "Python parent id": 7019, "Python id": 7020, "finished": true, "Ev Idx": 9920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908840.916, "dur": 0.349,
    "args": {
      "Python parent id": 7020, "Python id": 7021, "finished": true, "Ev Idx": 9921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908841.026, "dur": 0.192,
    "args": {
      "Python parent id": 7021, "Python id": 7022, "finished": true, "Ev Idx": 9922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908841.491, "dur": 0.505,
    "args": {
      "Python parent id": 7019, "Python id": 7023, "finished": true, "Ev Idx": 9923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908842.490, "dur": 0.350,
    "args": {
      "Python parent id": 6988, "Python id": 7024, "finished": true, "Ev Idx": 9924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908843.621, "dur": 1.559,
    "args": {
      "Python parent id": 6988, "Python id": 7025, "finished": true, "Ev Idx": 9925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908844.304, "dur": 0.768,
    "args": {
      "Python parent id": 7025, "Python id": 7026, "finished": true, "Ev Idx": 9926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908845.771, "dur": 4.143,
    "args": {
      "Python parent id": 6987, "Python id": 7027, "finished": true, "Ev Idx": 9927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908850.510, "dur": 10.550,
    "args": {
      "Python parent id": 12, "Python id": 7028, "finished": true, "Ev Idx": 9928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740908861.525, "dur": 0.168,
    "args": {
      "Python parent id": 12, "Python id": 7029, "finished": true, "Ev Idx": 9929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908862.426, "dur": 0.338,
    "args": {
      "Python parent id": 12, "Python id": 7030, "finished": true, "Ev Idx": 9930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908863.242, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 7031, "finished": true, "Ev Idx": 9931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740908863.673, "dur": 14.502,
    "args": {
      "Python parent id": 12, "Python id": 7032, "finished": true, "Ev Idx": 9932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908879.701, "dur": 16.287,
    "args": {
      "Python parent id": 12, "Python id": 7033, "finished": true, "Ev Idx": 9933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908880.088, "dur": 11.443,
    "args": {
      "Python parent id": 7033, "Python id": 7034, "finished": true, "Ev Idx": 9934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908880.248, "dur": 0.541,
    "args": {
      "Python parent id": 7034, "Python id": 7035, "finished": true, "Ev Idx": 9935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908880.419, "dur": 0.293,
    "args": {
      "Python parent id": 7035, "Python id": 7036, "finished": true, "Ev Idx": 9936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908880.525, "dur": 0.134,
    "args": {
      "Python parent id": 7036, "Python id": 7037, "finished": true, "Ev Idx": 9937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908881.093, "dur": 8.461,
    "args": {
      "Python parent id": 7034, "Python id": 7038, "finished": true, "Ev Idx": 9938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908881.196, "dur": 0.061,
    "args": {
      "Python parent id": 7038, "Python id": 7039, "finished": true, "Ev Idx": 9939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908881.379, "dur": 0.058,
    "args": {
      "Python parent id": 7038, "Python id": 7040, "finished": true, "Ev Idx": 9940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908881.573, "dur": 0.056,
    "args": {
      "Python parent id": 7038, "Python id": 7041, "finished": true, "Ev Idx": 9941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908882.036, "dur": 0.059,
    "args": {
      "Python parent id": 7038, "Python id": 7042, "finished": true, "Ev Idx": 9942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908882.365, "dur": 0.070,
    "args": {
      "Python parent id": 7038, "Python id": 7043, "finished": true, "Ev Idx": 9943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908882.760, "dur": 6.746,
    "args": {
      "Python parent id": 7038, "Python id": 7044, "finished": true, "Ev Idx": 9944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908882.835, "dur": 0.057,
    "args": {
      "Python parent id": 7044, "Python id": 7045, "finished": true, "Ev Idx": 9945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908883.058, "dur": 0.062,
    "args": {
      "Python parent id": 7044, "Python id": 7046, "finished": true, "Ev Idx": 9946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908883.290, "dur": 0.063,
    "args": {
      "Python parent id": 7044, "Python id": 7047, "finished": true, "Ev Idx": 9947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908883.658, "dur": 0.035,
    "args": {
      "Python parent id": 7044, "Python id": 7048, "finished": true, "Ev Idx": 9948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908883.952, "dur": 5.467,
    "args": {
      "Python parent id": 7044, "Python id": 7049, "finished": true, "Ev Idx": 9949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740908884.207, "dur": 5.095,
    "args": {
      "Python parent id": 7049, "Python id": 7050, "finished": true, "Ev Idx": 9950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740908884.321, "dur": 3.313,
    "args": {
      "Python parent id": 7050, "Python id": 7051, "finished": true, "Ev Idx": 9951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740908884.488, "dur": 3.107,
    "args": {
      "Python parent id": 7051, "Python id": 7052, "finished": true, "Ev Idx": 9952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740908884.636, "dur": 0.258,
    "args": {
      "Python parent id": 7052, "Python id": 7053, "finished": true, "Ev Idx": 9953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740908884.750, "dur": 0.060,
    "args": {
      "Python parent id": 7053, "Python id": 7054, "finished": true, "Ev Idx": 9954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.053, "dur": 2.225,
    "args": {
      "Python parent id": 7052, "Python id": 7055, "finished": true, "Ev Idx": 9955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.245, "dur": 1.915,
    "args": {
      "Python parent id": 7055, "Python id": 7056, "finished": true, "Ev Idx": 9956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.411, "dur": 1.705,
    "args": {
      "Python parent id": 7056, "Python id": 7057, "finished": true, "Ev Idx": 9957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.560, "dur": 1.266,
    "args": {
      "Python parent id": 7057, "Python id": 7058, "finished": true, "Ev Idx": 9958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.789, "dur": 0.502,
    "args": {
      "Python parent id": 7058, "Python id": 7059, "finished": true, "Ev Idx": 9959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908885.864, "dur": 0.049,
    "args": {
      "Python parent id": 7059, "Python id": 7060, "finished": true, "Ev Idx": 9960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740908886.063, "dur": 0.154,
    "args": {
      "Python parent id": 7059, "Python id": 7061, "finished": true, "Ev Idx": 9961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740908887.453, "dur": 0.104,
    "args": {
      "Python parent id": 7052, "Python id": 7062, "finished": true, "Ev Idx": 9962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740908887.818, "dur": 0.092,
    "args": {
      "Python parent id": 7050, "Python id": 7063, "finished": true, "Ev Idx": 9963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.160, "dur": 1.097,
    "args": {
      "Python parent id": 7050, "Python id": 7064, "finished": true, "Ev Idx": 9964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.299, "dur": 0.917,
    "args": {
      "Python parent id": 7064, "Python id": 7065, "finished": true, "Ev Idx": 9965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.438, "dur": 0.375,
    "args": {
      "Python parent id": 7065, "Python id": 7066, "finished": true, "Ev Idx": 9966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.574, "dur": 0.183,
    "args": {
      "Python parent id": 7066, "Python id": 7067, "finished": true, "Ev Idx": 9967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.671, "dur": 0.059,
    "args": {
      "Python parent id": 7067, "Python id": 7068, "finished": true, "Ev Idx": 9968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740908888.926, "dur": 0.258,
    "args": {
      "Python parent id": 7065, "Python id": 7069, "finished": true, "Ev Idx": 9969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740908889.673, "dur": 0.250,
    "args": {
      "Python parent id": 7034, "Python id": 7070, "finished": true, "Ev Idx": 9970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740908890.478, "dur": 0.928,
    "args": {
      "Python parent id": 7034, "Python id": 7071, "finished": true, "Ev Idx": 9971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740908890.853, "dur": 0.504,
    "args": {
      "Python parent id": 7071, "Python id": 7072, "finished": true, "Ev Idx": 9972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740908891.824, "dur": 4.089,
    "args": {
      "Python parent id": 7033, "Python id": 7073, "finished": true, "Ev Idx": 9973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740908896.891, "dur": 98.504,
    "args": {
      "Python parent id": 12, "Python id": 7074, "finished": true, "Ev Idx": 9974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908897.233, "dur": 0.567,
    "args": {
      "Python parent id": 7074, "Python id": 7075, "finished": true, "Ev Idx": 9975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908897.570, "dur": 0.144,
    "args": {
      "Python parent id": 7075, "Python id": 7076, "finished": true, "Ev Idx": 9976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908898.158, "dur": 0.578,
    "args": {
      "Python parent id": 7074, "Python id": 7077, "finished": true, "Ev Idx": 9977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740908898.929, "dur": 0.263,
    "args": {
      "Python parent id": 7074, "Python id": 7078, "finished": true, "Ev Idx": 9978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740908899.430, "dur": 1.828,
    "args": {
      "Python parent id": 7074, "Python id": 7079, "finished": true, "Ev Idx": 9979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740908901.818, "dur": 93.511,
    "args": {
      "Python parent id": 7074, "Python id": 7080, "finished": true, "Ev Idx": 9980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740908902.781, "dur": 92.344,
    "args": {
      "Python parent id": 7080, "Python id": 7081, "finished": true, "Ev Idx": 9981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908903.074, "dur": 91.997,
    "args": {
      "Python parent id": 7081, "Python id": 7082, "finished": true, "Ev Idx": 9982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740908909.468, "dur": 84.140,
    "args": {
      "Python parent id": 7082, "Python id": 7083, "finished": true, "Ev Idx": 9983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740908909.769, "dur": 83.787,
    "args": {
      "Python parent id": 7083, "Python id": 7084, "finished": true, "Ev Idx": 9984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740908910.308, "dur": 83.104,
    "args": {
      "Python parent id": 7084, "Python id": 7085, "finished": true, "Ev Idx": 9985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908910.766, "dur": 0.537,
    "args": {
      "Python parent id": 7085, "Python id": 7086, "finished": true, "Ev Idx": 9986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908911.178, "dur": 0.069,
    "args": {
      "Python parent id": 7086, "Python id": 7087, "finished": true, "Ev Idx": 9987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740908911.631, "dur": 0.316,
    "args": {
      "Python parent id": 7085, "Python id": 7088, "finished": true, "Ev Idx": 9988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740908911.844, "dur": 0.065,
    "args": {
      "Python parent id": 7088, "Python id": 7089, "finished": true, "Ev Idx": 9989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908912.258, "dur": 81.029,
    "args": {
      "Python parent id": 7085, "Python id": 7090, "finished": true, "Ev Idx": 9990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740908912.984, "dur": 79.027,
    "args": {
      "Python parent id": 7090, "Python id": 7091, "finished": true, "Ev Idx": 9991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740908920.601, "dur": 20.724,
    "args": {
      "Python parent id": 7091, "Python id": 7092, "finished": true, "Ev Idx": 9992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740908920.727, "dur": 0.122,
    "args": {
      "Python parent id": 7092, "Python id": 7093, "finished": true, "Ev Idx": 9993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740908921.108, "dur": 20.075,
    "args": {
      "Python parent id": 7092, "Python id": 7094, "finished": true, "Ev Idx": 9994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740908921.226, "dur": 9.558,
    "args": {
      "Python parent id": 7094, "Python id": 7095, "finished": true, "Ev Idx": 9995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908948.440, "dur": 8.137,
    "args": {
      "Python parent id": 7091, "Python id": 7096, "finished": true, "Ev Idx": 9996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740908967.201, "dur": 8.409,
    "args": {
      "Python parent id": 7091, "Python id": 7097, "finished": true, "Ev Idx": 9997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740908996.705, "dur": 17.493,
    "args": {
      "Python parent id": 12, "Python id": 7098, "finished": true, "Ev Idx": 9998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740908997.092, "dur": 12.960,
    "args": {
      "Python parent id": 7098, "Python id": 7099, "finished": true, "Ev Idx": 9999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740908997.266, "dur": 0.580,
    "args": {
      "Python parent id": 7099, "Python id": 7100, "finished": true, "Ev Idx": 10000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740908997.410, "dur": 0.335,
    "args": {
      "Python parent id": 7100, "Python id": 7101, "finished": true, "Ev Idx": 10001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740908997.572, "dur": 0.121,
    "args": {
      "Python parent id": 7101, "Python id": 7102, "finished": true, "Ev Idx": 10002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908998.207, "dur": 9.395,
    "args": {
      "Python parent id": 7099, "Python id": 7103, "finished": true, "Ev Idx": 10003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908998.361, "dur": 0.059,
    "args": {
      "Python parent id": 7103, "Python id": 7104, "finished": true, "Ev Idx": 10004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908998.523, "dur": 0.072,
    "args": {
      "Python parent id": 7103, "Python id": 7105, "finished": true, "Ev Idx": 10005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908998.741, "dur": 0.056,
    "args": {
      "Python parent id": 7103, "Python id": 7106, "finished": true, "Ev Idx": 10006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740908999.227, "dur": 0.072,
    "args": {
      "Python parent id": 7103, "Python id": 7107, "finished": true, "Ev Idx": 10007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740908999.529, "dur": 0.086,
    "args": {
      "Python parent id": 7103, "Python id": 7108, "finished": true, "Ev Idx": 10008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740908999.952, "dur": 7.577,
    "args": {
      "Python parent id": 7103, "Python id": 7109, "finished": true, "Ev Idx": 10009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909000.065, "dur": 0.061,
    "args": {
      "Python parent id": 7109, "Python id": 7110, "finished": true, "Ev Idx": 10010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909000.300, "dur": 0.056,
    "args": {
      "Python parent id": 7109, "Python id": 7111, "finished": true, "Ev Idx": 10011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909000.532, "dur": 0.066,
    "args": {
      "Python parent id": 7109, "Python id": 7112, "finished": true, "Ev Idx": 10012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909000.956, "dur": 0.042,
    "args": {
      "Python parent id": 7109, "Python id": 7113, "finished": true, "Ev Idx": 10013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909001.322, "dur": 6.088,
    "args": {
      "Python parent id": 7109, "Python id": 7114, "finished": true, "Ev Idx": 10014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909001.645, "dur": 5.648,
    "args": {
      "Python parent id": 7114, "Python id": 7115, "finished": true, "Ev Idx": 10015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909001.836, "dur": 3.713,
    "args": {
      "Python parent id": 7115, "Python id": 7116, "finished": true, "Ev Idx": 10016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909002.157, "dur": 3.298,
    "args": {
      "Python parent id": 7116, "Python id": 7117, "finished": true, "Ev Idx": 10017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909002.291, "dur": 0.260,
    "args": {
      "Python parent id": 7117, "Python id": 7118, "finished": true, "Ev Idx": 10018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909002.412, "dur": 0.061,
    "args": {
      "Python parent id": 7118, "Python id": 7119, "finished": true, "Ev Idx": 10019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909002.684, "dur": 2.362,
    "args": {
      "Python parent id": 7117, "Python id": 7120, "finished": true, "Ev Idx": 10020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909002.900, "dur": 2.036,
    "args": {
      "Python parent id": 7120, "Python id": 7121, "finished": true, "Ev Idx": 10021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909003.156, "dur": 1.739,
    "args": {
      "Python parent id": 7121, "Python id": 7122, "finished": true, "Ev Idx": 10022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909003.396, "dur": 1.262,
    "args": {
      "Python parent id": 7122, "Python id": 7123, "finished": true, "Ev Idx": 10023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909003.584, "dur": 0.585,
    "args": {
      "Python parent id": 7123, "Python id": 7124, "finished": true, "Ev Idx": 10024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909003.676, "dur": 0.045,
    "args": {
      "Python parent id": 7124, "Python id": 7125, "finished": true, "Ev Idx": 10025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909003.935, "dur": 0.167,
    "args": {
      "Python parent id": 7124, "Python id": 7126, "finished": true, "Ev Idx": 10026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909005.234, "dur": 0.163,
    "args": {
      "Python parent id": 7117, "Python id": 7127, "finished": true, "Ev Idx": 10027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909005.742, "dur": 0.110,
    "args": {
      "Python parent id": 7115, "Python id": 7128, "finished": true, "Ev Idx": 10028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.140, "dur": 1.096,
    "args": {
      "Python parent id": 7115, "Python id": 7129, "finished": true, "Ev Idx": 10029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.282, "dur": 0.915,
    "args": {
      "Python parent id": 7129, "Python id": 7130, "finished": true, "Ev Idx": 10030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.417, "dur": 0.372,
    "args": {
      "Python parent id": 7130, "Python id": 7131, "finished": true, "Ev Idx": 10031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.549, "dur": 0.171,
    "args": {
      "Python parent id": 7131, "Python id": 7132, "finished": true, "Ev Idx": 10032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.636, "dur": 0.055,
    "args": {
      "Python parent id": 7132, "Python id": 7133, "finished": true, "Ev Idx": 10033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909006.909, "dur": 0.260,
    "args": {
      "Python parent id": 7130, "Python id": 7134, "finished": true, "Ev Idx": 10034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909007.713, "dur": 0.302,
    "args": {
      "Python parent id": 7099, "Python id": 7135, "finished": true, "Ev Idx": 10035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909008.672, "dur": 1.225,
    "args": {
      "Python parent id": 7099, "Python id": 7136, "finished": true, "Ev Idx": 10036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909009.187, "dur": 0.645,
    "args": {
      "Python parent id": 7136, "Python id": 7137, "finished": true, "Ev Idx": 10037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909010.361, "dur": 3.745,
    "args": {
      "Python parent id": 7098, "Python id": 7138, "finished": true, "Ev Idx": 10038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909014.783, "dur": 8.320,
    "args": {
      "Python parent id": 12, "Python id": 7139, "finished": true, "Ev Idx": 10039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909023.546, "dur": 0.106,
    "args": {
      "Python parent id": 12, "Python id": 7140, "finished": true, "Ev Idx": 10040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909024.419, "dur": 0.201,
    "args": {
      "Python parent id": 12, "Python id": 7141, "finished": true, "Ev Idx": 10041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909025.091, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 7142, "finished": true, "Ev Idx": 10042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909025.526, "dur": 10.272,
    "args": {
      "Python parent id": 12, "Python id": 7143, "finished": true, "Ev Idx": 10043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909036.979, "dur": 14.571,
    "args": {
      "Python parent id": 12, "Python id": 7144, "finished": true, "Ev Idx": 10044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909037.373, "dur": 11.135,
    "args": {
      "Python parent id": 7144, "Python id": 7145, "finished": true, "Ev Idx": 10045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909037.528, "dur": 0.551,
    "args": {
      "Python parent id": 7145, "Python id": 7146, "finished": true, "Ev Idx": 10046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909037.679, "dur": 0.296,
    "args": {
      "Python parent id": 7146, "Python id": 7147, "finished": true, "Ev Idx": 10047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909037.818, "dur": 0.097,
    "args": {
      "Python parent id": 7147, "Python id": 7148, "finished": true, "Ev Idx": 10048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909038.364, "dur": 8.269,
    "args": {
      "Python parent id": 7145, "Python id": 7149, "finished": true, "Ev Idx": 10049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909038.501, "dur": 0.063,
    "args": {
      "Python parent id": 7149, "Python id": 7150, "finished": true, "Ev Idx": 10050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909038.731, "dur": 0.054,
    "args": {
      "Python parent id": 7149, "Python id": 7151, "finished": true, "Ev Idx": 10051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909038.934, "dur": 0.056,
    "args": {
      "Python parent id": 7149, "Python id": 7152, "finished": true, "Ev Idx": 10052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909039.317, "dur": 0.041,
    "args": {
      "Python parent id": 7149, "Python id": 7153, "finished": true, "Ev Idx": 10053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909039.637, "dur": 0.074,
    "args": {
      "Python parent id": 7149, "Python id": 7154, "finished": true, "Ev Idx": 10054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909040.078, "dur": 6.482,
    "args": {
      "Python parent id": 7149, "Python id": 7155, "finished": true, "Ev Idx": 10055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909040.141, "dur": 0.057,
    "args": {
      "Python parent id": 7155, "Python id": 7156, "finished": true, "Ev Idx": 10056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909040.352, "dur": 0.054,
    "args": {
      "Python parent id": 7155, "Python id": 7157, "finished": true, "Ev Idx": 10057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909040.562, "dur": 0.063,
    "args": {
      "Python parent id": 7155, "Python id": 7158, "finished": true, "Ev Idx": 10058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909040.978, "dur": 0.046,
    "args": {
      "Python parent id": 7155, "Python id": 7159, "finished": true, "Ev Idx": 10059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909041.377, "dur": 5.079,
    "args": {
      "Python parent id": 7155, "Python id": 7160, "finished": true, "Ev Idx": 10060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909041.612, "dur": 4.768,
    "args": {
      "Python parent id": 7160, "Python id": 7161, "finished": true, "Ev Idx": 10061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909041.746, "dur": 3.013,
    "args": {
      "Python parent id": 7161, "Python id": 7162, "finished": true, "Ev Idx": 10062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909041.929, "dur": 2.749,
    "args": {
      "Python parent id": 7162, "Python id": 7163, "finished": true, "Ev Idx": 10063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909042.057, "dur": 0.300,
    "args": {
      "Python parent id": 7163, "Python id": 7164, "finished": true, "Ev Idx": 10064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909042.202, "dur": 0.064,
    "args": {
      "Python parent id": 7164, "Python id": 7165, "finished": true, "Ev Idx": 10065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909042.525, "dur": 1.828,
    "args": {
      "Python parent id": 7163, "Python id": 7166, "finished": true, "Ev Idx": 10066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909042.719, "dur": 1.578,
    "args": {
      "Python parent id": 7166, "Python id": 7167, "finished": true, "Ev Idx": 10067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909042.898, "dur": 1.357,
    "args": {
      "Python parent id": 7167, "Python id": 7168, "finished": true, "Ev Idx": 10068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909043.046, "dur": 1.010,
    "args": {
      "Python parent id": 7168, "Python id": 7169, "finished": true, "Ev Idx": 10069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909043.200, "dur": 0.436,
    "args": {
      "Python parent id": 7169, "Python id": 7170, "finished": true, "Ev Idx": 10070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909043.288, "dur": 0.045,
    "args": {
      "Python parent id": 7170, "Python id": 7171, "finished": true, "Ev Idx": 10071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909043.455, "dur": 0.130,
    "args": {
      "Python parent id": 7170, "Python id": 7172, "finished": true, "Ev Idx": 10072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909044.521, "dur": 0.097,
    "args": {
      "Python parent id": 7163, "Python id": 7173, "finished": true, "Ev Idx": 10073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909044.922, "dur": 0.140,
    "args": {
      "Python parent id": 7161, "Python id": 7174, "finished": true, "Ev Idx": 10074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909045.315, "dur": 1.024,
    "args": {
      "Python parent id": 7161, "Python id": 7175, "finished": true, "Ev Idx": 10075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909045.440, "dur": 0.855,
    "args": {
      "Python parent id": 7175, "Python id": 7176, "finished": true, "Ev Idx": 10076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909045.566, "dur": 0.346,
    "args": {
      "Python parent id": 7176, "Python id": 7177, "finished": true, "Ev Idx": 10077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909045.692, "dur": 0.164,
    "args": {
      "Python parent id": 7177, "Python id": 7178, "finished": true, "Ev Idx": 10078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909045.773, "dur": 0.050,
    "args": {
      "Python parent id": 7178, "Python id": 7179, "finished": true, "Ev Idx": 10079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909046.040, "dur": 0.224,
    "args": {
      "Python parent id": 7176, "Python id": 7180, "finished": true, "Ev Idx": 10080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909046.736, "dur": 0.172,
    "args": {
      "Python parent id": 7145, "Python id": 7181, "finished": true, "Ev Idx": 10081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909047.490, "dur": 0.902,
    "args": {
      "Python parent id": 7145, "Python id": 7182, "finished": true, "Ev Idx": 10082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909047.877, "dur": 0.469,
    "args": {
      "Python parent id": 7182, "Python id": 7183, "finished": true, "Ev Idx": 10083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909048.796, "dur": 2.683,
    "args": {
      "Python parent id": 7144, "Python id": 7184, "finished": true, "Ev Idx": 10084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909052.425, "dur": 89.761,
    "args": {
      "Python parent id": 12, "Python id": 7185, "finished": true, "Ev Idx": 10085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909052.750, "dur": 1.236,
    "args": {
      "Python parent id": 7185, "Python id": 7186, "finished": true, "Ev Idx": 10086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909053.812, "dur": 0.113,
    "args": {
      "Python parent id": 7186, "Python id": 7187, "finished": true, "Ev Idx": 10087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909054.301, "dur": 0.424,
    "args": {
      "Python parent id": 7185, "Python id": 7188, "finished": true, "Ev Idx": 10088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909054.902, "dur": 0.281,
    "args": {
      "Python parent id": 7185, "Python id": 7189, "finished": true, "Ev Idx": 10089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909055.391, "dur": 1.746,
    "args": {
      "Python parent id": 7185, "Python id": 7190, "finished": true, "Ev Idx": 10090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909057.679, "dur": 84.413,
    "args": {
      "Python parent id": 7185, "Python id": 7191, "finished": true, "Ev Idx": 10091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909058.623, "dur": 83.257,
    "args": {
      "Python parent id": 7191, "Python id": 7192, "finished": true, "Ev Idx": 10092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909058.920, "dur": 82.900,
    "args": {
      "Python parent id": 7192, "Python id": 7193, "finished": true, "Ev Idx": 10093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909064.136, "dur": 76.322,
    "args": {
      "Python parent id": 7193, "Python id": 7194, "finished": true, "Ev Idx": 10094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909064.478, "dur": 75.907,
    "args": {
      "Python parent id": 7194, "Python id": 7195, "finished": true, "Ev Idx": 10095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909065.030, "dur": 75.214,
    "args": {
      "Python parent id": 7195, "Python id": 7196, "finished": true, "Ev Idx": 10096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909065.615, "dur": 0.388,
    "args": {
      "Python parent id": 7196, "Python id": 7197, "finished": true, "Ev Idx": 10097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909065.876, "dur": 0.083,
    "args": {
      "Python parent id": 7197, "Python id": 7198, "finished": true, "Ev Idx": 10098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909066.335, "dur": 0.333,
    "args": {
      "Python parent id": 7196, "Python id": 7199, "finished": true, "Ev Idx": 10099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909066.581, "dur": 0.037,
    "args": {
      "Python parent id": 7199, "Python id": 7200, "finished": true, "Ev Idx": 10100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909067.017, "dur": 73.115,
    "args": {
      "Python parent id": 7196, "Python id": 7201, "finished": true, "Ev Idx": 10101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909067.660, "dur": 71.226,
    "args": {
      "Python parent id": 7201, "Python id": 7202, "finished": true, "Ev Idx": 10102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909074.325, "dur": 18.832,
    "args": {
      "Python parent id": 7202, "Python id": 7203, "finished": true, "Ev Idx": 10103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909074.442, "dur": 0.125,
    "args": {
      "Python parent id": 7203, "Python id": 7204, "finished": true, "Ev Idx": 10104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909074.811, "dur": 18.166,
    "args": {
      "Python parent id": 7203, "Python id": 7205, "finished": true, "Ev Idx": 10105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909074.922, "dur": 8.846,
    "args": {
      "Python parent id": 7205, "Python id": 7206, "finished": true, "Ev Idx": 10106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909100.396, "dur": 6.961,
    "args": {
      "Python parent id": 7202, "Python id": 7207, "finished": true, "Ev Idx": 10107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909115.737, "dur": 8.355,
    "args": {
      "Python parent id": 7202, "Python id": 7208, "finished": true, "Ev Idx": 10108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909143.376, "dur": 16.313,
    "args": {
      "Python parent id": 12, "Python id": 7209, "finished": true, "Ev Idx": 10109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909143.698, "dur": 12.216,
    "args": {
      "Python parent id": 7209, "Python id": 7210, "finished": true, "Ev Idx": 10110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909143.957, "dur": 0.480,
    "args": {
      "Python parent id": 7210, "Python id": 7211, "finished": true, "Ev Idx": 10111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909144.126, "dur": 0.249,
    "args": {
      "Python parent id": 7211, "Python id": 7212, "finished": true, "Ev Idx": 10112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909144.233, "dur": 0.074,
    "args": {
      "Python parent id": 7212, "Python id": 7213, "finished": true, "Ev Idx": 10113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909144.784, "dur": 9.068,
    "args": {
      "Python parent id": 7210, "Python id": 7214, "finished": true, "Ev Idx": 10114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909144.942, "dur": 0.059,
    "args": {
      "Python parent id": 7214, "Python id": 7215, "finished": true, "Ev Idx": 10115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909145.126, "dur": 0.055,
    "args": {
      "Python parent id": 7214, "Python id": 7216, "finished": true, "Ev Idx": 10116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909145.338, "dur": 0.058,
    "args": {
      "Python parent id": 7214, "Python id": 7217, "finished": true, "Ev Idx": 10117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909145.842, "dur": 0.062,
    "args": {
      "Python parent id": 7214, "Python id": 7218, "finished": true, "Ev Idx": 10118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909146.162, "dur": 0.071,
    "args": {
      "Python parent id": 7214, "Python id": 7219, "finished": true, "Ev Idx": 10119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909146.610, "dur": 7.175,
    "args": {
      "Python parent id": 7214, "Python id": 7220, "finished": true, "Ev Idx": 10120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909146.732, "dur": 0.056,
    "args": {
      "Python parent id": 7220, "Python id": 7221, "finished": true, "Ev Idx": 10121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909146.957, "dur": 0.055,
    "args": {
      "Python parent id": 7220, "Python id": 7222, "finished": true, "Ev Idx": 10122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909147.166, "dur": 0.069,
    "args": {
      "Python parent id": 7220, "Python id": 7223, "finished": true, "Ev Idx": 10123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909147.611, "dur": 0.036,
    "args": {
      "Python parent id": 7220, "Python id": 7224, "finished": true, "Ev Idx": 10124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909147.930, "dur": 5.782,
    "args": {
      "Python parent id": 7220, "Python id": 7225, "finished": true, "Ev Idx": 10125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909148.158, "dur": 5.470,
    "args": {
      "Python parent id": 7225, "Python id": 7226, "finished": true, "Ev Idx": 10126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909148.281, "dur": 3.553,
    "args": {
      "Python parent id": 7226, "Python id": 7227, "finished": true, "Ev Idx": 10127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909148.494, "dur": 3.277,
    "args": {
      "Python parent id": 7227, "Python id": 7228, "finished": true, "Ev Idx": 10128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909148.609, "dur": 0.290,
    "args": {
      "Python parent id": 7228, "Python id": 7229, "finished": true, "Ev Idx": 10129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909148.753, "dur": 0.060,
    "args": {
      "Python parent id": 7229, "Python id": 7230, "finished": true, "Ev Idx": 10130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.077, "dur": 2.203,
    "args": {
      "Python parent id": 7228, "Python id": 7231, "finished": true, "Ev Idx": 10131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.268, "dur": 1.791,
    "args": {
      "Python parent id": 7231, "Python id": 7232, "finished": true, "Ev Idx": 10132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.435, "dur": 1.582,
    "args": {
      "Python parent id": 7232, "Python id": 7233, "finished": true, "Ev Idx": 10133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.635, "dur": 1.068,
    "args": {
      "Python parent id": 7233, "Python id": 7234, "finished": true, "Ev Idx": 10134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.805, "dur": 0.433,
    "args": {
      "Python parent id": 7234, "Python id": 7235, "finished": true, "Ev Idx": 10135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909149.881, "dur": 0.042,
    "args": {
      "Python parent id": 7235, "Python id": 7236, "finished": true, "Ev Idx": 10136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909150.052, "dur": 0.144,
    "args": {
      "Python parent id": 7235, "Python id": 7237, "finished": true, "Ev Idx": 10137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909151.544, "dur": 0.164,
    "args": {
      "Python parent id": 7228, "Python id": 7238, "finished": true, "Ev Idx": 10138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.006, "dur": 0.090,
    "args": {
      "Python parent id": 7226, "Python id": 7239, "finished": true, "Ev Idx": 10139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.422, "dur": 1.164,
    "args": {
      "Python parent id": 7226, "Python id": 7240, "finished": true, "Ev Idx": 10140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.588, "dur": 0.955,
    "args": {
      "Python parent id": 7240, "Python id": 7241, "finished": true, "Ev Idx": 10141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.713, "dur": 0.372,
    "args": {
      "Python parent id": 7241, "Python id": 7242, "finished": true, "Ev Idx": 10142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.840, "dur": 0.187,
    "args": {
      "Python parent id": 7242, "Python id": 7243, "finished": true, "Ev Idx": 10143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909152.941, "dur": 0.058,
    "args": {
      "Python parent id": 7243, "Python id": 7244, "finished": true, "Ev Idx": 10144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909153.214, "dur": 0.299,
    "args": {
      "Python parent id": 7241, "Python id": 7245, "finished": true, "Ev Idx": 10145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909153.973, "dur": 0.217,
    "args": {
      "Python parent id": 7210, "Python id": 7246, "finished": true, "Ev Idx": 10146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909154.846, "dur": 0.946,
    "args": {
      "Python parent id": 7210, "Python id": 7247, "finished": true, "Ev Idx": 10147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909155.239, "dur": 0.505,
    "args": {
      "Python parent id": 7247, "Python id": 7248, "finished": true, "Ev Idx": 10148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909156.205, "dur": 3.391,
    "args": {
      "Python parent id": 7209, "Python id": 7249, "finished": true, "Ev Idx": 10149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909160.254, "dur": 7.398,
    "args": {
      "Python parent id": 12, "Python id": 7250, "finished": true, "Ev Idx": 10150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909168.133, "dur": 0.119,
    "args": {
      "Python parent id": 12, "Python id": 7251, "finished": true, "Ev Idx": 10151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909168.819, "dur": 0.222,
    "args": {
      "Python parent id": 12, "Python id": 7252, "finished": true, "Ev Idx": 10152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909169.531, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 7253, "finished": true, "Ev Idx": 10153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909170.022, "dur": 9.758,
    "args": {
      "Python parent id": 12, "Python id": 7254, "finished": true, "Ev Idx": 10154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909181.023, "dur": 14.299,
    "args": {
      "Python parent id": 12, "Python id": 7255, "finished": true, "Ev Idx": 10155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909181.319, "dur": 10.834,
    "args": {
      "Python parent id": 7255, "Python id": 7256, "finished": true, "Ev Idx": 10156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909181.509, "dur": 0.444,
    "args": {
      "Python parent id": 7256, "Python id": 7257, "finished": true, "Ev Idx": 10157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909181.644, "dur": 0.233,
    "args": {
      "Python parent id": 7257, "Python id": 7258, "finished": true, "Ev Idx": 10158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909181.744, "dur": 0.099,
    "args": {
      "Python parent id": 7258, "Python id": 7259, "finished": true, "Ev Idx": 10159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909182.242, "dur": 8.157,
    "args": {
      "Python parent id": 7256, "Python id": 7260, "finished": true, "Ev Idx": 10160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909182.360, "dur": 0.071,
    "args": {
      "Python parent id": 7260, "Python id": 7261, "finished": true, "Ev Idx": 10161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909182.562, "dur": 0.055,
    "args": {
      "Python parent id": 7260, "Python id": 7262, "finished": true, "Ev Idx": 10162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909182.772, "dur": 0.062,
    "args": {
      "Python parent id": 7260, "Python id": 7263, "finished": true, "Ev Idx": 10163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909183.283, "dur": 0.034,
    "args": {
      "Python parent id": 7260, "Python id": 7264, "finished": true, "Ev Idx": 10164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909183.570, "dur": 0.073,
    "args": {
      "Python parent id": 7260, "Python id": 7265, "finished": true, "Ev Idx": 10165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909183.981, "dur": 6.349,
    "args": {
      "Python parent id": 7260, "Python id": 7266, "finished": true, "Ev Idx": 10166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909184.107, "dur": 0.057,
    "args": {
      "Python parent id": 7266, "Python id": 7267, "finished": true, "Ev Idx": 10167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909184.317, "dur": 0.060,
    "args": {
      "Python parent id": 7266, "Python id": 7268, "finished": true, "Ev Idx": 10168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909184.555, "dur": 0.069,
    "args": {
      "Python parent id": 7266, "Python id": 7269, "finished": true, "Ev Idx": 10169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909184.967, "dur": 0.032,
    "args": {
      "Python parent id": 7266, "Python id": 7270, "finished": true, "Ev Idx": 10170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909185.279, "dur": 4.952,
    "args": {
      "Python parent id": 7266, "Python id": 7271, "finished": true, "Ev Idx": 10171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909185.516, "dur": 4.659,
    "args": {
      "Python parent id": 7271, "Python id": 7272, "finished": true, "Ev Idx": 10172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909185.680, "dur": 2.838,
    "args": {
      "Python parent id": 7272, "Python id": 7273, "finished": true, "Ev Idx": 10173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909185.855, "dur": 2.619,
    "args": {
      "Python parent id": 7273, "Python id": 7274, "finished": true, "Ev Idx": 10174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909185.983, "dur": 0.276,
    "args": {
      "Python parent id": 7274, "Python id": 7275, "finished": true, "Ev Idx": 10175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909186.100, "dur": 0.062,
    "args": {
      "Python parent id": 7275, "Python id": 7276, "finished": true, "Ev Idx": 10176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909186.394, "dur": 1.759,
    "args": {
      "Python parent id": 7274, "Python id": 7277, "finished": true, "Ev Idx": 10177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909186.549, "dur": 1.546,
    "args": {
      "Python parent id": 7277, "Python id": 7278, "finished": true, "Ev Idx": 10178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909186.755, "dur": 1.294,
    "args": {
      "Python parent id": 7278, "Python id": 7279, "finished": true, "Ev Idx": 10179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909186.888, "dur": 0.985,
    "args": {
      "Python parent id": 7279, "Python id": 7280, "finished": true, "Ev Idx": 10180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909187.031, "dur": 0.451,
    "args": {
      "Python parent id": 7280, "Python id": 7281, "finished": true, "Ev Idx": 10181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909187.113, "dur": 0.043,
    "args": {
      "Python parent id": 7281, "Python id": 7282, "finished": true, "Ev Idx": 10182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909187.315, "dur": 0.113,
    "args": {
      "Python parent id": 7281, "Python id": 7283, "finished": true, "Ev Idx": 10183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909188.338, "dur": 0.085,
    "args": {
      "Python parent id": 7274, "Python id": 7284, "finished": true, "Ev Idx": 10184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909188.671, "dur": 0.103,
    "args": {
      "Python parent id": 7272, "Python id": 7285, "finished": true, "Ev Idx": 10185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.055, "dur": 1.078,
    "args": {
      "Python parent id": 7272, "Python id": 7286, "finished": true, "Ev Idx": 10186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.190, "dur": 0.900,
    "args": {
      "Python parent id": 7286, "Python id": 7287, "finished": true, "Ev Idx": 10187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.331, "dur": 0.375,
    "args": {
      "Python parent id": 7287, "Python id": 7288, "finished": true, "Ev Idx": 10188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.494, "dur": 0.156,
    "args": {
      "Python parent id": 7288, "Python id": 7289, "finished": true, "Ev Idx": 10189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.570, "dur": 0.047,
    "args": {
      "Python parent id": 7289, "Python id": 7290, "finished": true, "Ev Idx": 10190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909189.827, "dur": 0.237,
    "args": {
      "Python parent id": 7287, "Python id": 7291, "finished": true, "Ev Idx": 10191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909190.505, "dur": 0.172,
    "args": {
      "Python parent id": 7256, "Python id": 7292, "finished": true, "Ev Idx": 10192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909191.229, "dur": 0.811,
    "args": {
      "Python parent id": 7256, "Python id": 7293, "finished": true, "Ev Idx": 10193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909191.571, "dur": 0.420,
    "args": {
      "Python parent id": 7293, "Python id": 7294, "finished": true, "Ev Idx": 10194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909192.373, "dur": 2.876,
    "args": {
      "Python parent id": 7255, "Python id": 7295, "finished": true, "Ev Idx": 10195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909196.144, "dur": 89.059,
    "args": {
      "Python parent id": 12, "Python id": 7296, "finished": true, "Ev Idx": 10196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909196.459, "dur": 0.461,
    "args": {
      "Python parent id": 7296, "Python id": 7297, "finished": true, "Ev Idx": 10197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909196.738, "dur": 0.126,
    "args": {
      "Python parent id": 7297, "Python id": 7298, "finished": true, "Ev Idx": 10198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909197.261, "dur": 0.319,
    "args": {
      "Python parent id": 7296, "Python id": 7299, "finished": true, "Ev Idx": 10199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909197.776, "dur": 0.232,
    "args": {
      "Python parent id": 7296, "Python id": 7300, "finished": true, "Ev Idx": 10200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909198.213, "dur": 1.544,
    "args": {
      "Python parent id": 7296, "Python id": 7301, "finished": true, "Ev Idx": 10201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909200.314, "dur": 84.787,
    "args": {
      "Python parent id": 7296, "Python id": 7302, "finished": true, "Ev Idx": 10202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909201.143, "dur": 83.737,
    "args": {
      "Python parent id": 7302, "Python id": 7303, "finished": true, "Ev Idx": 10203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909201.407, "dur": 83.374,
    "args": {
      "Python parent id": 7303, "Python id": 7304, "finished": true, "Ev Idx": 10204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909206.451, "dur": 77.063,
    "args": {
      "Python parent id": 7304, "Python id": 7305, "finished": true, "Ev Idx": 10205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909206.777, "dur": 76.639,
    "args": {
      "Python parent id": 7305, "Python id": 7306, "finished": true, "Ev Idx": 10206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909207.291, "dur": 75.998,
    "args": {
      "Python parent id": 7306, "Python id": 7307, "finished": true, "Ev Idx": 10207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909207.792, "dur": 0.365,
    "args": {
      "Python parent id": 7307, "Python id": 7308, "finished": true, "Ev Idx": 10208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909208.036, "dur": 0.034,
    "args": {
      "Python parent id": 7308, "Python id": 7309, "finished": true, "Ev Idx": 10209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909208.397, "dur": 0.268,
    "args": {
      "Python parent id": 7307, "Python id": 7310, "finished": true, "Ev Idx": 10210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909208.578, "dur": 0.033,
    "args": {
      "Python parent id": 7310, "Python id": 7311, "finished": true, "Ev Idx": 10211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909208.970, "dur": 74.163,
    "args": {
      "Python parent id": 7307, "Python id": 7312, "finished": true, "Ev Idx": 10212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909209.678, "dur": 72.214,
    "args": {
      "Python parent id": 7312, "Python id": 7313, "finished": true, "Ev Idx": 10213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909216.448, "dur": 20.337,
    "args": {
      "Python parent id": 7313, "Python id": 7314, "finished": true, "Ev Idx": 10214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909216.572, "dur": 0.119,
    "args": {
      "Python parent id": 7314, "Python id": 7315, "finished": true, "Ev Idx": 10215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909216.931, "dur": 19.711,
    "args": {
      "Python parent id": 7314, "Python id": 7316, "finished": true, "Ev Idx": 10216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909217.048, "dur": 10.074,
    "args": {
      "Python parent id": 7316, "Python id": 7317, "finished": true, "Ev Idx": 10217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909244.418, "dur": 6.386,
    "args": {
      "Python parent id": 7313, "Python id": 7318, "finished": true, "Ev Idx": 10218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909259.289, "dur": 8.181,
    "args": {
      "Python parent id": 7313, "Python id": 7319, "finished": true, "Ev Idx": 10219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909286.333, "dur": 16.318,
    "args": {
      "Python parent id": 12, "Python id": 7320, "finished": true, "Ev Idx": 10220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909286.697, "dur": 11.911,
    "args": {
      "Python parent id": 7320, "Python id": 7321, "finished": true, "Ev Idx": 10221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909286.921, "dur": 0.464,
    "args": {
      "Python parent id": 7321, "Python id": 7322, "finished": true, "Ev Idx": 10222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909287.110, "dur": 0.212,
    "args": {
      "Python parent id": 7322, "Python id": 7323, "finished": true, "Ev Idx": 10223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909287.209, "dur": 0.064,
    "args": {
      "Python parent id": 7323, "Python id": 7324, "finished": true, "Ev Idx": 10224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909287.717, "dur": 8.764,
    "args": {
      "Python parent id": 7321, "Python id": 7325, "finished": true, "Ev Idx": 10225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909287.818, "dur": 0.075,
    "args": {
      "Python parent id": 7325, "Python id": 7326, "finished": true, "Ev Idx": 10226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909288.014, "dur": 0.071,
    "args": {
      "Python parent id": 7325, "Python id": 7327, "finished": true, "Ev Idx": 10227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909288.237, "dur": 0.054,
    "args": {
      "Python parent id": 7325, "Python id": 7328, "finished": true, "Ev Idx": 10228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909288.748, "dur": 0.056,
    "args": {
      "Python parent id": 7325, "Python id": 7329, "finished": true, "Ev Idx": 10229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909289.035, "dur": 0.071,
    "args": {
      "Python parent id": 7325, "Python id": 7330, "finished": true, "Ev Idx": 10230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909289.444, "dur": 6.956,
    "args": {
      "Python parent id": 7325, "Python id": 7331, "finished": true, "Ev Idx": 10231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909289.523, "dur": 0.061,
    "args": {
      "Python parent id": 7331, "Python id": 7332, "finished": true, "Ev Idx": 10232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909289.718, "dur": 0.062,
    "args": {
      "Python parent id": 7331, "Python id": 7333, "finished": true, "Ev Idx": 10233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909289.926, "dur": 0.067,
    "args": {
      "Python parent id": 7331, "Python id": 7334, "finished": true, "Ev Idx": 10234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909290.321, "dur": 0.065,
    "args": {
      "Python parent id": 7331, "Python id": 7335, "finished": true, "Ev Idx": 10235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909290.683, "dur": 5.674,
    "args": {
      "Python parent id": 7331, "Python id": 7336, "finished": true, "Ev Idx": 10236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909290.958, "dur": 5.301,
    "args": {
      "Python parent id": 7336, "Python id": 7337, "finished": true, "Ev Idx": 10237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909291.120, "dur": 3.371,
    "args": {
      "Python parent id": 7337, "Python id": 7338, "finished": true, "Ev Idx": 10238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909291.295, "dur": 3.128,
    "args": {
      "Python parent id": 7338, "Python id": 7339, "finished": true, "Ev Idx": 10239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909291.493, "dur": 0.240,
    "args": {
      "Python parent id": 7339, "Python id": 7340, "finished": true, "Ev Idx": 10240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909291.591, "dur": 0.068,
    "args": {
      "Python parent id": 7340, "Python id": 7341, "finished": true, "Ev Idx": 10241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909291.922, "dur": 2.132,
    "args": {
      "Python parent id": 7339, "Python id": 7342, "finished": true, "Ev Idx": 10242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.114, "dur": 1.831,
    "args": {
      "Python parent id": 7342, "Python id": 7343, "finished": true, "Ev Idx": 10243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.344, "dur": 1.511,
    "args": {
      "Python parent id": 7343, "Python id": 7344, "finished": true, "Ev Idx": 10244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.511, "dur": 1.128,
    "args": {
      "Python parent id": 7344, "Python id": 7345, "finished": true, "Ev Idx": 10245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.694, "dur": 0.475,
    "args": {
      "Python parent id": 7345, "Python id": 7346, "finished": true, "Ev Idx": 10246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.775, "dur": 0.049,
    "args": {
      "Python parent id": 7346, "Python id": 7347, "finished": true, "Ev Idx": 10247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909292.975, "dur": 0.154,
    "args": {
      "Python parent id": 7346, "Python id": 7348, "finished": true, "Ev Idx": 10248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909294.250, "dur": 0.137,
    "args": {
      "Python parent id": 7339, "Python id": 7349, "finished": true, "Ev Idx": 10249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909294.606, "dur": 0.097,
    "args": {
      "Python parent id": 7337, "Python id": 7350, "finished": true, "Ev Idx": 10250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.019, "dur": 1.183,
    "args": {
      "Python parent id": 7337, "Python id": 7351, "finished": true, "Ev Idx": 10251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.194, "dur": 0.957,
    "args": {
      "Python parent id": 7351, "Python id": 7352, "finished": true, "Ev Idx": 10252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.338, "dur": 0.388,
    "args": {
      "Python parent id": 7352, "Python id": 7353, "finished": true, "Ev Idx": 10253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.481, "dur": 0.172,
    "args": {
      "Python parent id": 7353, "Python id": 7354, "finished": true, "Ev Idx": 10254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.564, "dur": 0.058,
    "args": {
      "Python parent id": 7354, "Python id": 7355, "finished": true, "Ev Idx": 10255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909295.843, "dur": 0.276,
    "args": {
      "Python parent id": 7352, "Python id": 7356, "finished": true, "Ev Idx": 10256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909296.588, "dur": 0.265,
    "args": {
      "Python parent id": 7321, "Python id": 7357, "finished": true, "Ev Idx": 10257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909297.436, "dur": 1.030,
    "args": {
      "Python parent id": 7321, "Python id": 7358, "finished": true, "Ev Idx": 10258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909297.892, "dur": 0.527,
    "args": {
      "Python parent id": 7358, "Python id": 7359, "finished": true, "Ev Idx": 10259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909298.957, "dur": 3.631,
    "args": {
      "Python parent id": 7320, "Python id": 7360, "finished": true, "Ev Idx": 10260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909303.196, "dur": 7.407,
    "args": {
      "Python parent id": 12, "Python id": 7361, "finished": true, "Ev Idx": 10261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909311.058, "dur": 0.118,
    "args": {
      "Python parent id": 12, "Python id": 7362, "finished": true, "Ev Idx": 10262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909311.789, "dur": 0.238,
    "args": {
      "Python parent id": 12, "Python id": 7363, "finished": true, "Ev Idx": 10263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909312.495, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 7364, "finished": true, "Ev Idx": 10264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909312.955, "dur": 9.325,
    "args": {
      "Python parent id": 12, "Python id": 7365, "finished": true, "Ev Idx": 10265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909323.522, "dur": 14.422,
    "args": {
      "Python parent id": 12, "Python id": 7366, "finished": true, "Ev Idx": 10266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909323.867, "dur": 10.871,
    "args": {
      "Python parent id": 7366, "Python id": 7367, "finished": true, "Ev Idx": 10267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.006, "dur": 0.448,
    "args": {
      "Python parent id": 7367, "Python id": 7368, "finished": true, "Ev Idx": 10268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.162, "dur": 0.222,
    "args": {
      "Python parent id": 7368, "Python id": 7369, "finished": true, "Ev Idx": 10269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.243, "dur": 0.093,
    "args": {
      "Python parent id": 7369, "Python id": 7370, "finished": true, "Ev Idx": 10270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.684, "dur": 8.077,
    "args": {
      "Python parent id": 7367, "Python id": 7371, "finished": true, "Ev Idx": 10271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.764, "dur": 0.100,
    "args": {
      "Python parent id": 7371, "Python id": 7372, "finished": true, "Ev Idx": 10272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909324.979, "dur": 0.060,
    "args": {
      "Python parent id": 7371, "Python id": 7373, "finished": true, "Ev Idx": 10273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909325.173, "dur": 0.062,
    "args": {
      "Python parent id": 7371, "Python id": 7374, "finished": true, "Ev Idx": 10274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909325.665, "dur": 0.066,
    "args": {
      "Python parent id": 7371, "Python id": 7375, "finished": true, "Ev Idx": 10275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909325.987, "dur": 0.077,
    "args": {
      "Python parent id": 7371, "Python id": 7376, "finished": true, "Ev Idx": 10276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909326.365, "dur": 6.320,
    "args": {
      "Python parent id": 7371, "Python id": 7377, "finished": true, "Ev Idx": 10277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909326.456, "dur": 0.059,
    "args": {
      "Python parent id": 7377, "Python id": 7378, "finished": true, "Ev Idx": 10278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909326.682, "dur": 0.055,
    "args": {
      "Python parent id": 7377, "Python id": 7379, "finished": true, "Ev Idx": 10279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909326.910, "dur": 0.069,
    "args": {
      "Python parent id": 7377, "Python id": 7380, "finished": true, "Ev Idx": 10280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909327.294, "dur": 0.036,
    "args": {
      "Python parent id": 7377, "Python id": 7381, "finished": true, "Ev Idx": 10281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909327.645, "dur": 4.950,
    "args": {
      "Python parent id": 7377, "Python id": 7382, "finished": true, "Ev Idx": 10282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909327.903, "dur": 4.636,
    "args": {
      "Python parent id": 7382, "Python id": 7383, "finished": true, "Ev Idx": 10283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.023, "dur": 2.881,
    "args": {
      "Python parent id": 7383, "Python id": 7384, "finished": true, "Ev Idx": 10284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.199, "dur": 2.652,
    "args": {
      "Python parent id": 7384, "Python id": 7385, "finished": true, "Ev Idx": 10285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.331, "dur": 0.244,
    "args": {
      "Python parent id": 7385, "Python id": 7386, "finished": true, "Ev Idx": 10286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.436, "dur": 0.063,
    "args": {
      "Python parent id": 7386, "Python id": 7387, "finished": true, "Ev Idx": 10287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.763, "dur": 1.812,
    "args": {
      "Python parent id": 7385, "Python id": 7388, "finished": true, "Ev Idx": 10288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909328.918, "dur": 1.598,
    "args": {
      "Python parent id": 7388, "Python id": 7389, "finished": true, "Ev Idx": 10289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909329.099, "dur": 1.344,
    "args": {
      "Python parent id": 7389, "Python id": 7390, "finished": true, "Ev Idx": 10290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909329.219, "dur": 0.990,
    "args": {
      "Python parent id": 7390, "Python id": 7391, "finished": true, "Ev Idx": 10291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909329.376, "dur": 0.406,
    "args": {
      "Python parent id": 7391, "Python id": 7392, "finished": true, "Ev Idx": 10292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909329.457, "dur": 0.049,
    "args": {
      "Python parent id": 7392, "Python id": 7393, "finished": true, "Ev Idx": 10293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909329.628, "dur": 0.102,
    "args": {
      "Python parent id": 7392, "Python id": 7394, "finished": true, "Ev Idx": 10294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909330.746, "dur": 0.071,
    "args": {
      "Python parent id": 7385, "Python id": 7395, "finished": true, "Ev Idx": 10295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.068, "dur": 0.091,
    "args": {
      "Python parent id": 7383, "Python id": 7396, "finished": true, "Ev Idx": 10296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.427, "dur": 1.068,
    "args": {
      "Python parent id": 7383, "Python id": 7397, "finished": true, "Ev Idx": 10297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.559, "dur": 0.899,
    "args": {
      "Python parent id": 7397, "Python id": 7398, "finished": true, "Ev Idx": 10298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.700, "dur": 0.357,
    "args": {
      "Python parent id": 7398, "Python id": 7399, "finished": true, "Ev Idx": 10299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.826, "dur": 0.177,
    "args": {
      "Python parent id": 7399, "Python id": 7400, "finished": true, "Ev Idx": 10300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909331.914, "dur": 0.055,
    "args": {
      "Python parent id": 7400, "Python id": 7401, "finished": true, "Ev Idx": 10301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909332.180, "dur": 0.231,
    "args": {
      "Python parent id": 7398, "Python id": 7402, "finished": true, "Ev Idx": 10302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909332.860, "dur": 0.135,
    "args": {
      "Python parent id": 7367, "Python id": 7403, "finished": true, "Ev Idx": 10303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909333.626, "dur": 0.952,
    "args": {
      "Python parent id": 7367, "Python id": 7404, "finished": true, "Ev Idx": 10304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909334.074, "dur": 0.448,
    "args": {
      "Python parent id": 7404, "Python id": 7405, "finished": true, "Ev Idx": 10305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909335.023, "dur": 2.856,
    "args": {
      "Python parent id": 7366, "Python id": 7406, "finished": true, "Ev Idx": 10306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909338.796, "dur": 93.065,
    "args": {
      "Python parent id": 12, "Python id": 7407, "finished": true, "Ev Idx": 10307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909339.097, "dur": 0.425,
    "args": {
      "Python parent id": 7407, "Python id": 7408, "finished": true, "Ev Idx": 10308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909339.355, "dur": 0.077,
    "args": {
      "Python parent id": 7408, "Python id": 7409, "finished": true, "Ev Idx": 10309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909339.839, "dur": 0.353,
    "args": {
      "Python parent id": 7407, "Python id": 7410, "finished": true, "Ev Idx": 10310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909340.397, "dur": 0.245,
    "args": {
      "Python parent id": 7407, "Python id": 7411, "finished": true, "Ev Idx": 10311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909340.858, "dur": 1.405,
    "args": {
      "Python parent id": 7407, "Python id": 7412, "finished": true, "Ev Idx": 10312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909342.818, "dur": 88.950,
    "args": {
      "Python parent id": 7407, "Python id": 7413, "finished": true, "Ev Idx": 10313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909343.600, "dur": 87.965,
    "args": {
      "Python parent id": 7413, "Python id": 7414, "finished": true, "Ev Idx": 10314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909343.853, "dur": 87.640,
    "args": {
      "Python parent id": 7414, "Python id": 7415, "finished": true, "Ev Idx": 10315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909348.710, "dur": 81.489,
    "args": {
      "Python parent id": 7415, "Python id": 7416, "finished": true, "Ev Idx": 10316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909349.006, "dur": 81.095,
    "args": {
      "Python parent id": 7416, "Python id": 7417, "finished": true, "Ev Idx": 10317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909349.460, "dur": 80.497,
    "args": {
      "Python parent id": 7417, "Python id": 7418, "finished": true, "Ev Idx": 10318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909349.917, "dur": 0.377,
    "args": {
      "Python parent id": 7418, "Python id": 7419, "finished": true, "Ev Idx": 10319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909350.188, "dur": 0.035,
    "args": {
      "Python parent id": 7419, "Python id": 7420, "finished": true, "Ev Idx": 10320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909350.588, "dur": 0.306,
    "args": {
      "Python parent id": 7418, "Python id": 7421, "finished": true, "Ev Idx": 10321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909350.806, "dur": 0.035,
    "args": {
      "Python parent id": 7421, "Python id": 7422, "finished": true, "Ev Idx": 10322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909351.241, "dur": 78.591,
    "args": {
      "Python parent id": 7418, "Python id": 7423, "finished": true, "Ev Idx": 10323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909351.877, "dur": 76.810,
    "args": {
      "Python parent id": 7423, "Python id": 7424, "finished": true, "Ev Idx": 10324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909358.507, "dur": 18.146,
    "args": {
      "Python parent id": 7424, "Python id": 7425, "finished": true, "Ev Idx": 10325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909358.740, "dur": 0.134,
    "args": {
      "Python parent id": 7425, "Python id": 7426, "finished": true, "Ev Idx": 10326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909359.449, "dur": 17.023,
    "args": {
      "Python parent id": 7425, "Python id": 7427, "finished": true, "Ev Idx": 10327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909359.543, "dur": 7.754,
    "args": {
      "Python parent id": 7427, "Python id": 7428, "finished": true, "Ev Idx": 10328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909384.696, "dur": 6.989,
    "args": {
      "Python parent id": 7424, "Python id": 7429, "finished": true, "Ev Idx": 10329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909400.144, "dur": 7.564,
    "args": {
      "Python parent id": 7424, "Python id": 7430, "finished": true, "Ev Idx": 10330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909433.420, "dur": 17.281,
    "args": {
      "Python parent id": 12, "Python id": 7431, "finished": true, "Ev Idx": 10331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909433.785, "dur": 12.409,
    "args": {
      "Python parent id": 7431, "Python id": 7432, "finished": true, "Ev Idx": 10332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909433.991, "dur": 0.631,
    "args": {
      "Python parent id": 7432, "Python id": 7433, "finished": true, "Ev Idx": 10333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909434.175, "dur": 0.361,
    "args": {
      "Python parent id": 7433, "Python id": 7434, "finished": true, "Ev Idx": 10334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909434.328, "dur": 0.173,
    "args": {
      "Python parent id": 7434, "Python id": 7435, "finished": true, "Ev Idx": 10335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909434.985, "dur": 8.956,
    "args": {
      "Python parent id": 7432, "Python id": 7436, "finished": true, "Ev Idx": 10336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909435.131, "dur": 0.092,
    "args": {
      "Python parent id": 7436, "Python id": 7437, "finished": true, "Ev Idx": 10337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909435.347, "dur": 0.058,
    "args": {
      "Python parent id": 7436, "Python id": 7438, "finished": true, "Ev Idx": 10338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909435.535, "dur": 0.052,
    "args": {
      "Python parent id": 7436, "Python id": 7439, "finished": true, "Ev Idx": 10339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909436.022, "dur": 0.064,
    "args": {
      "Python parent id": 7436, "Python id": 7440, "finished": true, "Ev Idx": 10340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909436.354, "dur": 0.068,
    "args": {
      "Python parent id": 7436, "Python id": 7441, "finished": true, "Ev Idx": 10341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909436.779, "dur": 7.080,
    "args": {
      "Python parent id": 7436, "Python id": 7442, "finished": true, "Ev Idx": 10342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909436.886, "dur": 0.061,
    "args": {
      "Python parent id": 7442, "Python id": 7443, "finished": true, "Ev Idx": 10343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909437.113, "dur": 0.047,
    "args": {
      "Python parent id": 7442, "Python id": 7444, "finished": true, "Ev Idx": 10344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909437.323, "dur": 0.074,
    "args": {
      "Python parent id": 7442, "Python id": 7445, "finished": true, "Ev Idx": 10345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909437.809, "dur": 0.038,
    "args": {
      "Python parent id": 7442, "Python id": 7446, "finished": true, "Ev Idx": 10346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.125, "dur": 5.616,
    "args": {
      "Python parent id": 7442, "Python id": 7447, "finished": true, "Ev Idx": 10347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.353, "dur": 5.316,
    "args": {
      "Python parent id": 7447, "Python id": 7448, "finished": true, "Ev Idx": 10348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.478, "dur": 3.455,
    "args": {
      "Python parent id": 7448, "Python id": 7449, "finished": true, "Ev Idx": 10349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.661, "dur": 3.202,
    "args": {
      "Python parent id": 7449, "Python id": 7450, "finished": true, "Ev Idx": 10350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.785, "dur": 0.284,
    "args": {
      "Python parent id": 7450, "Python id": 7451, "finished": true, "Ev Idx": 10351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909438.913, "dur": 0.083,
    "args": {
      "Python parent id": 7451, "Python id": 7452, "finished": true, "Ev Idx": 10352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909439.234, "dur": 2.227,
    "args": {
      "Python parent id": 7450, "Python id": 7453, "finished": true, "Ev Idx": 10353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909439.468, "dur": 1.874,
    "args": {
      "Python parent id": 7453, "Python id": 7454, "finished": true, "Ev Idx": 10354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909439.670, "dur": 1.590,
    "args": {
      "Python parent id": 7454, "Python id": 7455, "finished": true, "Ev Idx": 10355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909439.821, "dur": 1.231,
    "args": {
      "Python parent id": 7455, "Python id": 7456, "finished": true, "Ev Idx": 10356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909439.987, "dur": 0.510,
    "args": {
      "Python parent id": 7456, "Python id": 7457, "finished": true, "Ev Idx": 10357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909440.065, "dur": 0.043,
    "args": {
      "Python parent id": 7457, "Python id": 7458, "finished": true, "Ev Idx": 10358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909440.268, "dur": 0.158,
    "args": {
      "Python parent id": 7457, "Python id": 7459, "finished": true, "Ev Idx": 10359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909441.642, "dur": 0.164,
    "args": {
      "Python parent id": 7450, "Python id": 7460, "finished": true, "Ev Idx": 10360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.128, "dur": 0.090,
    "args": {
      "Python parent id": 7448, "Python id": 7461, "finished": true, "Ev Idx": 10361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.467, "dur": 1.138,
    "args": {
      "Python parent id": 7448, "Python id": 7462, "finished": true, "Ev Idx": 10362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.607, "dur": 0.941,
    "args": {
      "Python parent id": 7462, "Python id": 7463, "finished": true, "Ev Idx": 10363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.728, "dur": 0.379,
    "args": {
      "Python parent id": 7463, "Python id": 7464, "finished": true, "Ev Idx": 10364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.869, "dur": 0.166,
    "args": {
      "Python parent id": 7464, "Python id": 7465, "finished": true, "Ev Idx": 10365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909442.954, "dur": 0.050,
    "args": {
      "Python parent id": 7465, "Python id": 7466, "finished": true, "Ev Idx": 10366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909443.257, "dur": 0.252,
    "args": {
      "Python parent id": 7463, "Python id": 7467, "finished": true, "Ev Idx": 10367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909444.048, "dur": 0.287,
    "args": {
      "Python parent id": 7432, "Python id": 7468, "finished": true, "Ev Idx": 10368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909444.965, "dur": 1.093,
    "args": {
      "Python parent id": 7432, "Python id": 7469, "finished": true, "Ev Idx": 10369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909445.442, "dur": 0.569,
    "args": {
      "Python parent id": 7469, "Python id": 7470, "finished": true, "Ev Idx": 10370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909446.489, "dur": 4.122,
    "args": {
      "Python parent id": 7431, "Python id": 7471, "finished": true, "Ev Idx": 10371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909451.215, "dur": 8.821,
    "args": {
      "Python parent id": 12, "Python id": 7472, "finished": true, "Ev Idx": 10372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909460.526, "dur": 0.135,
    "args": {
      "Python parent id": 12, "Python id": 7473, "finished": true, "Ev Idx": 10373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909461.347, "dur": 0.249,
    "args": {
      "Python parent id": 12, "Python id": 7474, "finished": true, "Ev Idx": 10374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909462.106, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 7475, "finished": true, "Ev Idx": 10375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909462.564, "dur": 10.033,
    "args": {
      "Python parent id": 12, "Python id": 7476, "finished": true, "Ev Idx": 10376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909473.841, "dur": 15.101,
    "args": {
      "Python parent id": 12, "Python id": 7477, "finished": true, "Ev Idx": 10377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909474.173, "dur": 10.964,
    "args": {
      "Python parent id": 7477, "Python id": 7478, "finished": true, "Ev Idx": 10378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909474.350, "dur": 0.421,
    "args": {
      "Python parent id": 7478, "Python id": 7479, "finished": true, "Ev Idx": 10379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909474.492, "dur": 0.212,
    "args": {
      "Python parent id": 7479, "Python id": 7480, "finished": true, "Ev Idx": 10380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909474.581, "dur": 0.091,
    "args": {
      "Python parent id": 7480, "Python id": 7481, "finished": true, "Ev Idx": 10381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909475.050, "dur": 8.188,
    "args": {
      "Python parent id": 7478, "Python id": 7482, "finished": true, "Ev Idx": 10382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909475.136, "dur": 0.099,
    "args": {
      "Python parent id": 7482, "Python id": 7483, "finished": true, "Ev Idx": 10383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909475.336, "dur": 0.059,
    "args": {
      "Python parent id": 7482, "Python id": 7484, "finished": true, "Ev Idx": 10384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909475.536, "dur": 0.057,
    "args": {
      "Python parent id": 7482, "Python id": 7485, "finished": true, "Ev Idx": 10385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909476.019, "dur": 0.037,
    "args": {
      "Python parent id": 7482, "Python id": 7486, "finished": true, "Ev Idx": 10386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909476.290, "dur": 0.071,
    "args": {
      "Python parent id": 7482, "Python id": 7487, "finished": true, "Ev Idx": 10387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909476.634, "dur": 6.488,
    "args": {
      "Python parent id": 7482, "Python id": 7488, "finished": true, "Ev Idx": 10388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909476.723, "dur": 0.057,
    "args": {
      "Python parent id": 7488, "Python id": 7489, "finished": true, "Ev Idx": 10389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909476.931, "dur": 0.050,
    "args": {
      "Python parent id": 7488, "Python id": 7490, "finished": true, "Ev Idx": 10390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909477.131, "dur": 0.065,
    "args": {
      "Python parent id": 7488, "Python id": 7491, "finished": true, "Ev Idx": 10391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909477.529, "dur": 0.034,
    "args": {
      "Python parent id": 7488, "Python id": 7492, "finished": true, "Ev Idx": 10392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909477.844, "dur": 5.221,
    "args": {
      "Python parent id": 7488, "Python id": 7493, "finished": true, "Ev Idx": 10393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909478.122, "dur": 4.891,
    "args": {
      "Python parent id": 7493, "Python id": 7494, "finished": true, "Ev Idx": 10394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909478.286, "dur": 3.085,
    "args": {
      "Python parent id": 7494, "Python id": 7495, "finished": true, "Ev Idx": 10395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909478.475, "dur": 2.812,
    "args": {
      "Python parent id": 7495, "Python id": 7496, "finished": true, "Ev Idx": 10396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909478.615, "dur": 0.275,
    "args": {
      "Python parent id": 7496, "Python id": 7497, "finished": true, "Ev Idx": 10397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909478.747, "dur": 0.062,
    "args": {
      "Python parent id": 7497, "Python id": 7498, "finished": true, "Ev Idx": 10398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.024, "dur": 1.954,
    "args": {
      "Python parent id": 7496, "Python id": 7499, "finished": true, "Ev Idx": 10399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.224, "dur": 1.672,
    "args": {
      "Python parent id": 7499, "Python id": 7500, "finished": true, "Ev Idx": 10400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.408, "dur": 1.444,
    "args": {
      "Python parent id": 7500, "Python id": 7501, "finished": true, "Ev Idx": 10401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.566, "dur": 1.092,
    "args": {
      "Python parent id": 7501, "Python id": 7502, "finished": true, "Ev Idx": 10402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.739, "dur": 0.440,
    "args": {
      "Python parent id": 7502, "Python id": 7503, "finished": true, "Ev Idx": 10403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.814, "dur": 0.045,
    "args": {
      "Python parent id": 7503, "Python id": 7504, "finished": true, "Ev Idx": 10404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909479.999, "dur": 0.122,
    "args": {
      "Python parent id": 7503, "Python id": 7505, "finished": true, "Ev Idx": 10405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909481.159, "dur": 0.067,
    "args": {
      "Python parent id": 7496, "Python id": 7506, "finished": true, "Ev Idx": 10406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909481.527, "dur": 0.087,
    "args": {
      "Python parent id": 7494, "Python id": 7507, "finished": true, "Ev Idx": 10407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909481.862, "dur": 1.092,
    "args": {
      "Python parent id": 7494, "Python id": 7508, "finished": true, "Ev Idx": 10408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909482.010, "dur": 0.908,
    "args": {
      "Python parent id": 7508, "Python id": 7509, "finished": true, "Ev Idx": 10409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909482.161, "dur": 0.384,
    "args": {
      "Python parent id": 7509, "Python id": 7510, "finished": true, "Ev Idx": 10410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909482.322, "dur": 0.150,
    "args": {
      "Python parent id": 7510, "Python id": 7511, "finished": true, "Ev Idx": 10411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909482.394, "dur": 0.050,
    "args": {
      "Python parent id": 7511, "Python id": 7512, "finished": true, "Ev Idx": 10412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909482.659, "dur": 0.211,
    "args": {
      "Python parent id": 7509, "Python id": 7513, "finished": true, "Ev Idx": 10413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909483.343, "dur": 0.233,
    "args": {
      "Python parent id": 7478, "Python id": 7514, "finished": true, "Ev Idx": 10414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909484.141, "dur": 0.880,
    "args": {
      "Python parent id": 7478, "Python id": 7515, "finished": true, "Ev Idx": 10415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909484.565, "dur": 0.410,
    "args": {
      "Python parent id": 7515, "Python id": 7516, "finished": true, "Ev Idx": 10416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909485.382, "dur": 3.491,
    "args": {
      "Python parent id": 7477, "Python id": 7517, "finished": true, "Ev Idx": 10417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909489.788, "dur": 88.829,
    "args": {
      "Python parent id": 12, "Python id": 7518, "finished": true, "Ev Idx": 10418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909490.199, "dur": 0.481,
    "args": {
      "Python parent id": 7518, "Python id": 7519, "finished": true, "Ev Idx": 10419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909490.484, "dur": 0.090,
    "args": {
      "Python parent id": 7519, "Python id": 7520, "finished": true, "Ev Idx": 10420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909490.998, "dur": 0.422,
    "args": {
      "Python parent id": 7518, "Python id": 7521, "finished": true, "Ev Idx": 10421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909491.591, "dur": 0.225,
    "args": {
      "Python parent id": 7518, "Python id": 7522, "finished": true, "Ev Idx": 10422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909492.044, "dur": 2.114,
    "args": {
      "Python parent id": 7518, "Python id": 7523, "finished": true, "Ev Idx": 10423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909494.749, "dur": 83.775,
    "args": {
      "Python parent id": 7518, "Python id": 7524, "finished": true, "Ev Idx": 10424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909495.704, "dur": 82.625,
    "args": {
      "Python parent id": 7524, "Python id": 7525, "finished": true, "Ev Idx": 10425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909495.947, "dur": 82.301,
    "args": {
      "Python parent id": 7525, "Python id": 7526, "finished": true, "Ev Idx": 10426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909500.886, "dur": 76.107,
    "args": {
      "Python parent id": 7526, "Python id": 7527, "finished": true, "Ev Idx": 10427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909501.213, "dur": 75.718,
    "args": {
      "Python parent id": 7527, "Python id": 7528, "finished": true, "Ev Idx": 10428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909501.673, "dur": 75.137,
    "args": {
      "Python parent id": 7528, "Python id": 7529, "finished": true, "Ev Idx": 10429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909502.196, "dur": 0.400,
    "args": {
      "Python parent id": 7529, "Python id": 7530, "finished": true, "Ev Idx": 10430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909502.490, "dur": 0.042,
    "args": {
      "Python parent id": 7530, "Python id": 7531, "finished": true, "Ev Idx": 10431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909502.913, "dur": 0.288,
    "args": {
      "Python parent id": 7529, "Python id": 7532, "finished": true, "Ev Idx": 10432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909503.107, "dur": 0.040,
    "args": {
      "Python parent id": 7532, "Python id": 7533, "finished": true, "Ev Idx": 10433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909503.539, "dur": 73.178,
    "args": {
      "Python parent id": 7529, "Python id": 7534, "finished": true, "Ev Idx": 10434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909504.115, "dur": 71.442,
    "args": {
      "Python parent id": 7534, "Python id": 7535, "finished": true, "Ev Idx": 10435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909510.917, "dur": 17.779,
    "args": {
      "Python parent id": 7535, "Python id": 7536, "finished": true, "Ev Idx": 10436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909511.027, "dur": 0.130,
    "args": {
      "Python parent id": 7536, "Python id": 7537, "finished": true, "Ev Idx": 10437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909511.397, "dur": 17.148,
    "args": {
      "Python parent id": 7536, "Python id": 7538, "finished": true, "Ev Idx": 10438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909511.495, "dur": 7.867,
    "args": {
      "Python parent id": 7538, "Python id": 7539, "finished": true, "Ev Idx": 10439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909535.882, "dur": 6.635,
    "args": {
      "Python parent id": 7535, "Python id": 7540, "finished": true, "Ev Idx": 10440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909551.790, "dur": 8.971,
    "args": {
      "Python parent id": 7535, "Python id": 7541, "finished": true, "Ev Idx": 10441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909579.747, "dur": 17.129,
    "args": {
      "Python parent id": 12, "Python id": 7542, "finished": true, "Ev Idx": 10442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909580.067, "dur": 12.422,
    "args": {
      "Python parent id": 7542, "Python id": 7543, "finished": true, "Ev Idx": 10443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909580.248, "dur": 0.484,
    "args": {
      "Python parent id": 7543, "Python id": 7544, "finished": true, "Ev Idx": 10444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909580.430, "dur": 0.239,
    "args": {
      "Python parent id": 7544, "Python id": 7545, "finished": true, "Ev Idx": 10445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909580.553, "dur": 0.082,
    "args": {
      "Python parent id": 7545, "Python id": 7546, "finished": true, "Ev Idx": 10446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909581.098, "dur": 9.139,
    "args": {
      "Python parent id": 7543, "Python id": 7547, "finished": true, "Ev Idx": 10447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909581.264, "dur": 0.085,
    "args": {
      "Python parent id": 7547, "Python id": 7548, "finished": true, "Ev Idx": 10448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909581.474, "dur": 0.058,
    "args": {
      "Python parent id": 7547, "Python id": 7549, "finished": true, "Ev Idx": 10449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909581.670, "dur": 0.053,
    "args": {
      "Python parent id": 7547, "Python id": 7550, "finished": true, "Ev Idx": 10450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909582.228, "dur": 0.034,
    "args": {
      "Python parent id": 7547, "Python id": 7551, "finished": true, "Ev Idx": 10451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909582.529, "dur": 0.081,
    "args": {
      "Python parent id": 7547, "Python id": 7552, "finished": true, "Ev Idx": 10452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909582.965, "dur": 7.181,
    "args": {
      "Python parent id": 7547, "Python id": 7553, "finished": true, "Ev Idx": 10453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909583.073, "dur": 0.077,
    "args": {
      "Python parent id": 7553, "Python id": 7554, "finished": true, "Ev Idx": 10454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909583.299, "dur": 0.058,
    "args": {
      "Python parent id": 7553, "Python id": 7555, "finished": true, "Ev Idx": 10455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909583.561, "dur": 0.064,
    "args": {
      "Python parent id": 7553, "Python id": 7556, "finished": true, "Ev Idx": 10456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909583.989, "dur": 0.054,
    "args": {
      "Python parent id": 7553, "Python id": 7557, "finished": true, "Ev Idx": 10457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909584.310, "dur": 5.737,
    "args": {
      "Python parent id": 7553, "Python id": 7558, "finished": true, "Ev Idx": 10458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909584.600, "dur": 5.374,
    "args": {
      "Python parent id": 7558, "Python id": 7559, "finished": true, "Ev Idx": 10459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909584.733, "dur": 3.366,
    "args": {
      "Python parent id": 7559, "Python id": 7560, "finished": true, "Ev Idx": 10460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909584.892, "dur": 3.135,
    "args": {
      "Python parent id": 7560, "Python id": 7561, "finished": true, "Ev Idx": 10461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909585.013, "dur": 0.289,
    "args": {
      "Python parent id": 7561, "Python id": 7562, "finished": true, "Ev Idx": 10462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909585.145, "dur": 0.082,
    "args": {
      "Python parent id": 7562, "Python id": 7563, "finished": true, "Ev Idx": 10463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909585.484, "dur": 2.164,
    "args": {
      "Python parent id": 7561, "Python id": 7564, "finished": true, "Ev Idx": 10464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909585.702, "dur": 1.846,
    "args": {
      "Python parent id": 7564, "Python id": 7565, "finished": true, "Ev Idx": 10465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909585.935, "dur": 1.566,
    "args": {
      "Python parent id": 7565, "Python id": 7566, "finished": true, "Ev Idx": 10466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909586.132, "dur": 1.142,
    "args": {
      "Python parent id": 7566, "Python id": 7567, "finished": true, "Ev Idx": 10467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909586.284, "dur": 0.550,
    "args": {
      "Python parent id": 7567, "Python id": 7568, "finished": true, "Ev Idx": 10468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909586.387, "dur": 0.045,
    "args": {
      "Python parent id": 7568, "Python id": 7569, "finished": true, "Ev Idx": 10469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909586.605, "dur": 0.167,
    "args": {
      "Python parent id": 7568, "Python id": 7570, "finished": true, "Ev Idx": 10470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909587.833, "dur": 0.136,
    "args": {
      "Python parent id": 7561, "Python id": 7571, "finished": true, "Ev Idx": 10471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909588.280, "dur": 0.104,
    "args": {
      "Python parent id": 7559, "Python id": 7572, "finished": true, "Ev Idx": 10472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909588.667, "dur": 1.247,
    "args": {
      "Python parent id": 7559, "Python id": 7573, "finished": true, "Ev Idx": 10473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909588.844, "dur": 1.033,
    "args": {
      "Python parent id": 7573, "Python id": 7574, "finished": true, "Ev Idx": 10474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909588.991, "dur": 0.374,
    "args": {
      "Python parent id": 7574, "Python id": 7575, "finished": true, "Ev Idx": 10475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909589.120, "dur": 0.174,
    "args": {
      "Python parent id": 7575, "Python id": 7576, "finished": true, "Ev Idx": 10476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909589.210, "dur": 0.056,
    "args": {
      "Python parent id": 7576, "Python id": 7577, "finished": true, "Ev Idx": 10477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909589.511, "dur": 0.304,
    "args": {
      "Python parent id": 7574, "Python id": 7578, "finished": true, "Ev Idx": 10478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909590.361, "dur": 0.276,
    "args": {
      "Python parent id": 7543, "Python id": 7579, "finished": true, "Ev Idx": 10479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909591.305, "dur": 1.061,
    "args": {
      "Python parent id": 7543, "Python id": 7580, "finished": true, "Ev Idx": 10480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909591.814, "dur": 0.504,
    "args": {
      "Python parent id": 7580, "Python id": 7581, "finished": true, "Ev Idx": 10481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909592.847, "dur": 3.951,
    "args": {
      "Python parent id": 7542, "Python id": 7582, "finished": true, "Ev Idx": 10482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909597.394, "dur": 7.721,
    "args": {
      "Python parent id": 12, "Python id": 7583, "finished": true, "Ev Idx": 10483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909605.599, "dur": 0.093,
    "args": {
      "Python parent id": 12, "Python id": 7584, "finished": true, "Ev Idx": 10484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909606.283, "dur": 0.180,
    "args": {
      "Python parent id": 12, "Python id": 7585, "finished": true, "Ev Idx": 10485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909606.938, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 7586, "finished": true, "Ev Idx": 10486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909607.421, "dur": 9.534,
    "args": {
      "Python parent id": 12, "Python id": 7587, "finished": true, "Ev Idx": 10487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909618.228, "dur": 14.595,
    "args": {
      "Python parent id": 12, "Python id": 7588, "finished": true, "Ev Idx": 10488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909618.520, "dur": 10.890,
    "args": {
      "Python parent id": 7588, "Python id": 7589, "finished": true, "Ev Idx": 10489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909618.675, "dur": 0.515,
    "args": {
      "Python parent id": 7589, "Python id": 7590, "finished": true, "Ev Idx": 10490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909618.837, "dur": 0.255,
    "args": {
      "Python parent id": 7590, "Python id": 7591, "finished": true, "Ev Idx": 10491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909618.929, "dur": 0.119,
    "args": {
      "Python parent id": 7591, "Python id": 7592, "finished": true, "Ev Idx": 10492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909619.453, "dur": 8.169,
    "args": {
      "Python parent id": 7589, "Python id": 7593, "finished": true, "Ev Idx": 10493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909619.561, "dur": 0.059,
    "args": {
      "Python parent id": 7593, "Python id": 7594, "finished": true, "Ev Idx": 10494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909619.722, "dur": 0.067,
    "args": {
      "Python parent id": 7593, "Python id": 7595, "finished": true, "Ev Idx": 10495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909619.918, "dur": 0.057,
    "args": {
      "Python parent id": 7593, "Python id": 7596, "finished": true, "Ev Idx": 10496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909620.406, "dur": 0.041,
    "args": {
      "Python parent id": 7593, "Python id": 7597, "finished": true, "Ev Idx": 10497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909620.705, "dur": 0.068,
    "args": {
      "Python parent id": 7593, "Python id": 7598, "finished": true, "Ev Idx": 10498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909621.150, "dur": 6.368,
    "args": {
      "Python parent id": 7593, "Python id": 7599, "finished": true, "Ev Idx": 10499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909621.241, "dur": 0.063,
    "args": {
      "Python parent id": 7599, "Python id": 7600, "finished": true, "Ev Idx": 10500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909621.452, "dur": 0.055,
    "args": {
      "Python parent id": 7599, "Python id": 7601, "finished": true, "Ev Idx": 10501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909621.692, "dur": 0.065,
    "args": {
      "Python parent id": 7599, "Python id": 7602, "finished": true, "Ev Idx": 10502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909622.130, "dur": 0.065,
    "args": {
      "Python parent id": 7599, "Python id": 7603, "finished": true, "Ev Idx": 10503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909622.455, "dur": 4.975,
    "args": {
      "Python parent id": 7599, "Python id": 7604, "finished": true, "Ev Idx": 10504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909622.663, "dur": 4.689,
    "args": {
      "Python parent id": 7604, "Python id": 7605, "finished": true, "Ev Idx": 10505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909622.793, "dur": 2.946,
    "args": {
      "Python parent id": 7605, "Python id": 7606, "finished": true, "Ev Idx": 10506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909622.984, "dur": 2.717,
    "args": {
      "Python parent id": 7606, "Python id": 7607, "finished": true, "Ev Idx": 10507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909623.122, "dur": 0.258,
    "args": {
      "Python parent id": 7607, "Python id": 7608, "finished": true, "Ev Idx": 10508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909623.236, "dur": 0.062,
    "args": {
      "Python parent id": 7608, "Python id": 7609, "finished": true, "Ev Idx": 10509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909623.533, "dur": 1.837,
    "args": {
      "Python parent id": 7607, "Python id": 7610, "finished": true, "Ev Idx": 10510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909623.684, "dur": 1.607,
    "args": {
      "Python parent id": 7610, "Python id": 7611, "finished": true, "Ev Idx": 10511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909623.872, "dur": 1.376,
    "args": {
      "Python parent id": 7611, "Python id": 7612, "finished": true, "Ev Idx": 10512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909624.008, "dur": 1.027,
    "args": {
      "Python parent id": 7612, "Python id": 7613, "finished": true, "Ev Idx": 10513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909624.159, "dur": 0.464,
    "args": {
      "Python parent id": 7613, "Python id": 7614, "finished": true, "Ev Idx": 10514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909624.239, "dur": 0.045,
    "args": {
      "Python parent id": 7614, "Python id": 7615, "finished": true, "Ev Idx": 10515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909624.434, "dur": 0.127,
    "args": {
      "Python parent id": 7614, "Python id": 7616, "finished": true, "Ev Idx": 10516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909625.554, "dur": 0.085,
    "args": {
      "Python parent id": 7607, "Python id": 7617, "finished": true, "Ev Idx": 10517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909625.881, "dur": 0.089,
    "args": {
      "Python parent id": 7605, "Python id": 7618, "finished": true, "Ev Idx": 10518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909626.242, "dur": 1.069,
    "args": {
      "Python parent id": 7605, "Python id": 7619, "finished": true, "Ev Idx": 10519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909626.389, "dur": 0.882,
    "args": {
      "Python parent id": 7619, "Python id": 7620, "finished": true, "Ev Idx": 10520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909626.559, "dur": 0.337,
    "args": {
      "Python parent id": 7620, "Python id": 7621, "finished": true, "Ev Idx": 10521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909626.672, "dur": 0.168,
    "args": {
      "Python parent id": 7621, "Python id": 7622, "finished": true, "Ev Idx": 10522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909626.759, "dur": 0.050,
    "args": {
      "Python parent id": 7622, "Python id": 7623, "finished": true, "Ev Idx": 10523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909627.014, "dur": 0.226,
    "args": {
      "Python parent id": 7620, "Python id": 7624, "finished": true, "Ev Idx": 10524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909627.733, "dur": 0.200,
    "args": {
      "Python parent id": 7589, "Python id": 7625, "finished": true, "Ev Idx": 10525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909628.425, "dur": 0.875,
    "args": {
      "Python parent id": 7589, "Python id": 7626, "finished": true, "Ev Idx": 10526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909628.820, "dur": 0.408,
    "args": {
      "Python parent id": 7626, "Python id": 7627, "finished": true, "Ev Idx": 10527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909629.676, "dur": 3.083,
    "args": {
      "Python parent id": 7588, "Python id": 7628, "finished": true, "Ev Idx": 10528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909633.667, "dur": 84.308,
    "args": {
      "Python parent id": 12, "Python id": 7629, "finished": true, "Ev Idx": 10529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909634.050, "dur": 0.411,
    "args": {
      "Python parent id": 7629, "Python id": 7630, "finished": true, "Ev Idx": 10530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909634.295, "dur": 0.109,
    "args": {
      "Python parent id": 7630, "Python id": 7631, "finished": true, "Ev Idx": 10531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909634.737, "dur": 0.364,
    "args": {
      "Python parent id": 7629, "Python id": 7632, "finished": true, "Ev Idx": 10532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909635.326, "dur": 0.233,
    "args": {
      "Python parent id": 7629, "Python id": 7633, "finished": true, "Ev Idx": 10533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909635.760, "dur": 1.540,
    "args": {
      "Python parent id": 7629, "Python id": 7634, "finished": true, "Ev Idx": 10534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909637.833, "dur": 80.076,
    "args": {
      "Python parent id": 7629, "Python id": 7635, "finished": true, "Ev Idx": 10535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909638.814, "dur": 78.929,
    "args": {
      "Python parent id": 7635, "Python id": 7636, "finished": true, "Ev Idx": 10536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909639.093, "dur": 78.594,
    "args": {
      "Python parent id": 7636, "Python id": 7637, "finished": true, "Ev Idx": 10537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909643.786, "dur": 72.693,
    "args": {
      "Python parent id": 7637, "Python id": 7638, "finished": true, "Ev Idx": 10538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909644.053, "dur": 72.351,
    "args": {
      "Python parent id": 7638, "Python id": 7639, "finished": true, "Ev Idx": 10539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909644.558, "dur": 71.706,
    "args": {
      "Python parent id": 7639, "Python id": 7640, "finished": true, "Ev Idx": 10540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909645.094, "dur": 0.392,
    "args": {
      "Python parent id": 7640, "Python id": 7641, "finished": true, "Ev Idx": 10541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909645.364, "dur": 0.047,
    "args": {
      "Python parent id": 7641, "Python id": 7642, "finished": true, "Ev Idx": 10542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909645.789, "dur": 0.293,
    "args": {
      "Python parent id": 7640, "Python id": 7643, "finished": true, "Ev Idx": 10543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909645.994, "dur": 0.039,
    "args": {
      "Python parent id": 7643, "Python id": 7644, "finished": true, "Ev Idx": 10544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909646.378, "dur": 69.748,
    "args": {
      "Python parent id": 7640, "Python id": 7645, "finished": true, "Ev Idx": 10545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909646.992, "dur": 68.004,
    "args": {
      "Python parent id": 7645, "Python id": 7646, "finished": true, "Ev Idx": 10546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909653.576, "dur": 17.097,
    "args": {
      "Python parent id": 7646, "Python id": 7647, "finished": true, "Ev Idx": 10547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909653.704, "dur": 0.122,
    "args": {
      "Python parent id": 7647, "Python id": 7648, "finished": true, "Ev Idx": 10548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909654.064, "dur": 16.449,
    "args": {
      "Python parent id": 7647, "Python id": 7649, "finished": true, "Ev Idx": 10549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909654.159, "dur": 7.047,
    "args": {
      "Python parent id": 7649, "Python id": 7650, "finished": true, "Ev Idx": 10550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909678.078, "dur": 6.493,
    "args": {
      "Python parent id": 7646, "Python id": 7651, "finished": true, "Ev Idx": 10551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909692.853, "dur": 7.370,
    "args": {
      "Python parent id": 7646, "Python id": 7652, "finished": true, "Ev Idx": 10552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909719.106, "dur": 15.896,
    "args": {
      "Python parent id": 12, "Python id": 7653, "finished": true, "Ev Idx": 10553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909719.482, "dur": 11.712,
    "args": {
      "Python parent id": 7653, "Python id": 7654, "finished": true, "Ev Idx": 10554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909719.677, "dur": 0.452,
    "args": {
      "Python parent id": 7654, "Python id": 7655, "finished": true, "Ev Idx": 10555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909719.861, "dur": 0.180,
    "args": {
      "Python parent id": 7655, "Python id": 7656, "finished": true, "Ev Idx": 10556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909719.936, "dur": 0.065,
    "args": {
      "Python parent id": 7656, "Python id": 7657, "finished": true, "Ev Idx": 10557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909720.443, "dur": 8.695,
    "args": {
      "Python parent id": 7654, "Python id": 7658, "finished": true, "Ev Idx": 10558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909720.542, "dur": 0.097,
    "args": {
      "Python parent id": 7658, "Python id": 7659, "finished": true, "Ev Idx": 10559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909720.773, "dur": 0.059,
    "args": {
      "Python parent id": 7658, "Python id": 7660, "finished": true, "Ev Idx": 10560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909721.010, "dur": 0.066,
    "args": {
      "Python parent id": 7658, "Python id": 7661, "finished": true, "Ev Idx": 10561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909721.519, "dur": 0.031,
    "args": {
      "Python parent id": 7658, "Python id": 7662, "finished": true, "Ev Idx": 10562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909721.806, "dur": 0.082,
    "args": {
      "Python parent id": 7658, "Python id": 7663, "finished": true, "Ev Idx": 10563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909722.230, "dur": 6.818,
    "args": {
      "Python parent id": 7658, "Python id": 7664, "finished": true, "Ev Idx": 10564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909722.362, "dur": 0.058,
    "args": {
      "Python parent id": 7664, "Python id": 7665, "finished": true, "Ev Idx": 10565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909722.566, "dur": 0.055,
    "args": {
      "Python parent id": 7664, "Python id": 7666, "finished": true, "Ev Idx": 10566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909722.791, "dur": 0.082,
    "args": {
      "Python parent id": 7664, "Python id": 7667, "finished": true, "Ev Idx": 10567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909723.238, "dur": 0.032,
    "args": {
      "Python parent id": 7664, "Python id": 7668, "finished": true, "Ev Idx": 10568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909723.593, "dur": 5.404,
    "args": {
      "Python parent id": 7664, "Python id": 7669, "finished": true, "Ev Idx": 10569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909723.912, "dur": 5.015,
    "args": {
      "Python parent id": 7669, "Python id": 7670, "finished": true, "Ev Idx": 10570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909724.036, "dur": 3.251,
    "args": {
      "Python parent id": 7670, "Python id": 7671, "finished": true, "Ev Idx": 10571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909724.232, "dur": 2.995,
    "args": {
      "Python parent id": 7671, "Python id": 7672, "finished": true, "Ev Idx": 10572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909724.388, "dur": 0.277,
    "args": {
      "Python parent id": 7672, "Python id": 7673, "finished": true, "Ev Idx": 10573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909724.507, "dur": 0.080,
    "args": {
      "Python parent id": 7673, "Python id": 7674, "finished": true, "Ev Idx": 10574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909724.832, "dur": 2.022,
    "args": {
      "Python parent id": 7672, "Python id": 7675, "finished": true, "Ev Idx": 10575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.040, "dur": 1.716,
    "args": {
      "Python parent id": 7675, "Python id": 7676, "finished": true, "Ev Idx": 10576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.233, "dur": 1.475,
    "args": {
      "Python parent id": 7676, "Python id": 7677, "finished": true, "Ev Idx": 10577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.390, "dur": 1.101,
    "args": {
      "Python parent id": 7677, "Python id": 7678, "finished": true, "Ev Idx": 10578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.549, "dur": 0.492,
    "args": {
      "Python parent id": 7678, "Python id": 7679, "finished": true, "Ev Idx": 10579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.633, "dur": 0.049,
    "args": {
      "Python parent id": 7679, "Python id": 7680, "finished": true, "Ev Idx": 10580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909725.821, "dur": 0.168,
    "args": {
      "Python parent id": 7679, "Python id": 7681, "finished": true, "Ev Idx": 10581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909727.026, "dur": 0.146,
    "args": {
      "Python parent id": 7672, "Python id": 7682, "finished": true, "Ev Idx": 10582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909727.419, "dur": 0.097,
    "args": {
      "Python parent id": 7670, "Python id": 7683, "finished": true, "Ev Idx": 10583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909727.776, "dur": 1.092,
    "args": {
      "Python parent id": 7670, "Python id": 7684, "finished": true, "Ev Idx": 10584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909727.948, "dur": 0.879,
    "args": {
      "Python parent id": 7684, "Python id": 7685, "finished": true, "Ev Idx": 10585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909728.086, "dur": 0.344,
    "args": {
      "Python parent id": 7685, "Python id": 7686, "finished": true, "Ev Idx": 10586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909728.212, "dur": 0.162,
    "args": {
      "Python parent id": 7686, "Python id": 7687, "finished": true, "Ev Idx": 10587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909728.294, "dur": 0.050,
    "args": {
      "Python parent id": 7687, "Python id": 7688, "finished": true, "Ev Idx": 10588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909728.547, "dur": 0.249,
    "args": {
      "Python parent id": 7685, "Python id": 7689, "finished": true, "Ev Idx": 10589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909729.241, "dur": 0.257,
    "args": {
      "Python parent id": 7654, "Python id": 7690, "finished": true, "Ev Idx": 10590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909730.060, "dur": 1.010,
    "args": {
      "Python parent id": 7654, "Python id": 7691, "finished": true, "Ev Idx": 10591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909730.506, "dur": 0.519,
    "args": {
      "Python parent id": 7691, "Python id": 7692, "finished": true, "Ev Idx": 10592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909731.455, "dur": 3.461,
    "args": {
      "Python parent id": 7653, "Python id": 7693, "finished": true, "Ev Idx": 10593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909735.495, "dur": 7.295,
    "args": {
      "Python parent id": 12, "Python id": 7694, "finished": true, "Ev Idx": 10594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909743.236, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 7695, "finished": true, "Ev Idx": 10595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909743.913, "dur": 0.132,
    "args": {
      "Python parent id": 12, "Python id": 7696, "finished": true, "Ev Idx": 10596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909744.485, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 7697, "finished": true, "Ev Idx": 10597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909744.929, "dur": 9.536,
    "args": {
      "Python parent id": 12, "Python id": 7698, "finished": true, "Ev Idx": 10598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909755.786, "dur": 14.083,
    "args": {
      "Python parent id": 12, "Python id": 7699, "finished": true, "Ev Idx": 10599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909756.133, "dur": 10.261,
    "args": {
      "Python parent id": 7699, "Python id": 7700, "finished": true, "Ev Idx": 10600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909756.315, "dur": 0.447,
    "args": {
      "Python parent id": 7700, "Python id": 7701, "finished": true, "Ev Idx": 10601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909756.460, "dur": 0.231,
    "args": {
      "Python parent id": 7701, "Python id": 7702, "finished": true, "Ev Idx": 10602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909756.561, "dur": 0.098,
    "args": {
      "Python parent id": 7702, "Python id": 7703, "finished": true, "Ev Idx": 10603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909757.058, "dur": 7.639,
    "args": {
      "Python parent id": 7700, "Python id": 7704, "finished": true, "Ev Idx": 10604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909757.171, "dur": 0.065,
    "args": {
      "Python parent id": 7704, "Python id": 7705, "finished": true, "Ev Idx": 10605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909757.339, "dur": 0.058,
    "args": {
      "Python parent id": 7704, "Python id": 7706, "finished": true, "Ev Idx": 10606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909757.543, "dur": 0.059,
    "args": {
      "Python parent id": 7704, "Python id": 7707, "finished": true, "Ev Idx": 10607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909758.012, "dur": 0.040,
    "args": {
      "Python parent id": 7704, "Python id": 7708, "finished": true, "Ev Idx": 10608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909758.301, "dur": 0.069,
    "args": {
      "Python parent id": 7704, "Python id": 7709, "finished": true, "Ev Idx": 10609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909758.662, "dur": 5.957,
    "args": {
      "Python parent id": 7704, "Python id": 7710, "finished": true, "Ev Idx": 10610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909758.750, "dur": 0.060,
    "args": {
      "Python parent id": 7710, "Python id": 7711, "finished": true, "Ev Idx": 10611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909758.952, "dur": 0.052,
    "args": {
      "Python parent id": 7710, "Python id": 7712, "finished": true, "Ev Idx": 10612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909759.151, "dur": 0.061,
    "args": {
      "Python parent id": 7710, "Python id": 7713, "finished": true, "Ev Idx": 10613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909759.475, "dur": 0.031,
    "args": {
      "Python parent id": 7710, "Python id": 7714, "finished": true, "Ev Idx": 10614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909759.833, "dur": 4.673,
    "args": {
      "Python parent id": 7710, "Python id": 7715, "finished": true, "Ev Idx": 10615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.032, "dur": 4.416,
    "args": {
      "Python parent id": 7715, "Python id": 7716, "finished": true, "Ev Idx": 10616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.174, "dur": 2.650,
    "args": {
      "Python parent id": 7716, "Python id": 7717, "finished": true, "Ev Idx": 10617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.338, "dur": 2.442,
    "args": {
      "Python parent id": 7717, "Python id": 7718, "finished": true, "Ev Idx": 10618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.470, "dur": 0.240,
    "args": {
      "Python parent id": 7718, "Python id": 7719, "finished": true, "Ev Idx": 10619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.599, "dur": 0.061,
    "args": {
      "Python parent id": 7719, "Python id": 7720, "finished": true, "Ev Idx": 10620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909760.883, "dur": 1.633,
    "args": {
      "Python parent id": 7718, "Python id": 7721, "finished": true, "Ev Idx": 10621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.037, "dur": 1.420,
    "args": {
      "Python parent id": 7721, "Python id": 7722, "finished": true, "Ev Idx": 10622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.212, "dur": 1.200,
    "args": {
      "Python parent id": 7722, "Python id": 7723, "finished": true, "Ev Idx": 10623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.327, "dur": 0.918,
    "args": {
      "Python parent id": 7723, "Python id": 7724, "finished": true, "Ev Idx": 10624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.494, "dur": 0.402,
    "args": {
      "Python parent id": 7724, "Python id": 7725, "finished": true, "Ev Idx": 10625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.566, "dur": 0.046,
    "args": {
      "Python parent id": 7725, "Python id": 7726, "finished": true, "Ev Idx": 10626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909761.734, "dur": 0.120,
    "args": {
      "Python parent id": 7725, "Python id": 7727, "finished": true, "Ev Idx": 10627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909762.677, "dur": 0.070,
    "args": {
      "Python parent id": 7718, "Python id": 7728, "finished": true, "Ev Idx": 10628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909762.988, "dur": 0.094,
    "args": {
      "Python parent id": 7716, "Python id": 7729, "finished": true, "Ev Idx": 10629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909763.329, "dur": 1.074,
    "args": {
      "Python parent id": 7716, "Python id": 7730, "finished": true, "Ev Idx": 10630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909763.465, "dur": 0.896,
    "args": {
      "Python parent id": 7730, "Python id": 7731, "finished": true, "Ev Idx": 10631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909763.599, "dur": 0.350,
    "args": {
      "Python parent id": 7731, "Python id": 7732, "finished": true, "Ev Idx": 10632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909763.722, "dur": 0.169,
    "args": {
      "Python parent id": 7732, "Python id": 7733, "finished": true, "Ev Idx": 10633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909763.805, "dur": 0.054,
    "args": {
      "Python parent id": 7733, "Python id": 7734, "finished": true, "Ev Idx": 10634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909764.097, "dur": 0.230,
    "args": {
      "Python parent id": 7731, "Python id": 7735, "finished": true, "Ev Idx": 10635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909764.802, "dur": 0.157,
    "args": {
      "Python parent id": 7700, "Python id": 7736, "finished": true, "Ev Idx": 10636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909765.480, "dur": 0.802,
    "args": {
      "Python parent id": 7700, "Python id": 7737, "finished": true, "Ev Idx": 10637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909765.842, "dur": 0.393,
    "args": {
      "Python parent id": 7737, "Python id": 7738, "finished": true, "Ev Idx": 10638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909766.695, "dur": 3.092,
    "args": {
      "Python parent id": 7699, "Python id": 7739, "finished": true, "Ev Idx": 10639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909770.724, "dur": 85.908,
    "args": {
      "Python parent id": 12, "Python id": 7740, "finished": true, "Ev Idx": 10640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909771.005, "dur": 0.447,
    "args": {
      "Python parent id": 7740, "Python id": 7741, "finished": true, "Ev Idx": 10641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909771.274, "dur": 0.120,
    "args": {
      "Python parent id": 7741, "Python id": 7742, "finished": true, "Ev Idx": 10642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909771.744, "dur": 0.338,
    "args": {
      "Python parent id": 7740, "Python id": 7743, "finished": true, "Ev Idx": 10643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909772.290, "dur": 0.266,
    "args": {
      "Python parent id": 7740, "Python id": 7744, "finished": true, "Ev Idx": 10644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909772.747, "dur": 1.529,
    "args": {
      "Python parent id": 7740, "Python id": 7745, "finished": true, "Ev Idx": 10645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909774.856, "dur": 81.710,
    "args": {
      "Python parent id": 7740, "Python id": 7746, "finished": true, "Ev Idx": 10646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909775.667, "dur": 80.691,
    "args": {
      "Python parent id": 7746, "Python id": 7747, "finished": true, "Ev Idx": 10647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909775.973, "dur": 80.310,
    "args": {
      "Python parent id": 7747, "Python id": 7748, "finished": true, "Ev Idx": 10648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909780.819, "dur": 74.188,
    "args": {
      "Python parent id": 7748, "Python id": 7749, "finished": true, "Ev Idx": 10649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909781.093, "dur": 73.852,
    "args": {
      "Python parent id": 7749, "Python id": 7750, "finished": true, "Ev Idx": 10650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909781.579, "dur": 73.265,
    "args": {
      "Python parent id": 7750, "Python id": 7751, "finished": true, "Ev Idx": 10651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909782.077, "dur": 0.360,
    "args": {
      "Python parent id": 7751, "Python id": 7752, "finished": true, "Ev Idx": 10652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909782.306, "dur": 0.089,
    "args": {
      "Python parent id": 7752, "Python id": 7753, "finished": true, "Ev Idx": 10653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909782.771, "dur": 0.328,
    "args": {
      "Python parent id": 7751, "Python id": 7754, "finished": true, "Ev Idx": 10654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909783.011, "dur": 0.035,
    "args": {
      "Python parent id": 7754, "Python id": 7755, "finished": true, "Ev Idx": 10655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909783.406, "dur": 71.375,
    "args": {
      "Python parent id": 7751, "Python id": 7756, "finished": true, "Ev Idx": 10656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909783.980, "dur": 69.584,
    "args": {
      "Python parent id": 7756, "Python id": 7757, "finished": true, "Ev Idx": 10657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909790.511, "dur": 17.483,
    "args": {
      "Python parent id": 7757, "Python id": 7758, "finished": true, "Ev Idx": 10658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909790.652, "dur": 0.124,
    "args": {
      "Python parent id": 7758, "Python id": 7759, "finished": true, "Ev Idx": 10659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909791.009, "dur": 16.881,
    "args": {
      "Python parent id": 7758, "Python id": 7760, "finished": true, "Ev Idx": 10660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909791.107, "dur": 7.286,
    "args": {
      "Python parent id": 7760, "Python id": 7761, "finished": true, "Ev Idx": 10661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909815.475, "dur": 6.958,
    "args": {
      "Python parent id": 7757, "Python id": 7762, "finished": true, "Ev Idx": 10662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909831.409, "dur": 7.303,
    "args": {
      "Python parent id": 7757, "Python id": 7763, "finished": true, "Ev Idx": 10663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909857.889, "dur": 16.043,
    "args": {
      "Python parent id": 12, "Python id": 7764, "finished": true, "Ev Idx": 10664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909858.187, "dur": 11.916,
    "args": {
      "Python parent id": 7764, "Python id": 7765, "finished": true, "Ev Idx": 10665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909858.419, "dur": 0.395,
    "args": {
      "Python parent id": 7765, "Python id": 7766, "finished": true, "Ev Idx": 10666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909858.562, "dur": 0.191,
    "args": {
      "Python parent id": 7766, "Python id": 7767, "finished": true, "Ev Idx": 10667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909858.652, "dur": 0.061,
    "args": {
      "Python parent id": 7767, "Python id": 7768, "finished": true, "Ev Idx": 10668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909859.109, "dur": 8.889,
    "args": {
      "Python parent id": 7765, "Python id": 7769, "finished": true, "Ev Idx": 10669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909859.225, "dur": 0.093,
    "args": {
      "Python parent id": 7769, "Python id": 7770, "finished": true, "Ev Idx": 10670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909859.418, "dur": 0.056,
    "args": {
      "Python parent id": 7769, "Python id": 7771, "finished": true, "Ev Idx": 10671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909859.656, "dur": 0.051,
    "args": {
      "Python parent id": 7769, "Python id": 7772, "finished": true, "Ev Idx": 10672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909860.150, "dur": 0.052,
    "args": {
      "Python parent id": 7769, "Python id": 7773, "finished": true, "Ev Idx": 10673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909860.477, "dur": 0.076,
    "args": {
      "Python parent id": 7769, "Python id": 7774, "finished": true, "Ev Idx": 10674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909860.905, "dur": 7.020,
    "args": {
      "Python parent id": 7769, "Python id": 7775, "finished": true, "Ev Idx": 10675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909861.037, "dur": 0.059,
    "args": {
      "Python parent id": 7775, "Python id": 7776, "finished": true, "Ev Idx": 10676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909861.230, "dur": 0.062,
    "args": {
      "Python parent id": 7775, "Python id": 7777, "finished": true, "Ev Idx": 10677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909861.488, "dur": 0.083,
    "args": {
      "Python parent id": 7775, "Python id": 7778, "finished": true, "Ev Idx": 10678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909861.896, "dur": 0.037,
    "args": {
      "Python parent id": 7775, "Python id": 7779, "finished": true, "Ev Idx": 10679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909862.258, "dur": 5.617,
    "args": {
      "Python parent id": 7775, "Python id": 7780, "finished": true, "Ev Idx": 10680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909862.533, "dur": 5.267,
    "args": {
      "Python parent id": 7780, "Python id": 7781, "finished": true, "Ev Idx": 10681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909862.678, "dur": 3.386,
    "args": {
      "Python parent id": 7781, "Python id": 7782, "finished": true, "Ev Idx": 10682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909862.891, "dur": 3.101,
    "args": {
      "Python parent id": 7782, "Python id": 7783, "finished": true, "Ev Idx": 10683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909863.054, "dur": 0.322,
    "args": {
      "Python parent id": 7783, "Python id": 7784, "finished": true, "Ev Idx": 10684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909863.243, "dur": 0.082,
    "args": {
      "Python parent id": 7784, "Python id": 7785, "finished": true, "Ev Idx": 10685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909863.534, "dur": 2.083,
    "args": {
      "Python parent id": 7783, "Python id": 7786, "finished": true, "Ev Idx": 10686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909863.759, "dur": 1.770,
    "args": {
      "Python parent id": 7786, "Python id": 7787, "finished": true, "Ev Idx": 10687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909863.974, "dur": 1.489,
    "args": {
      "Python parent id": 7787, "Python id": 7788, "finished": true, "Ev Idx": 10688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909864.154, "dur": 1.104,
    "args": {
      "Python parent id": 7788, "Python id": 7789, "finished": true, "Ev Idx": 10689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909864.323, "dur": 0.500,
    "args": {
      "Python parent id": 7789, "Python id": 7790, "finished": true, "Ev Idx": 10690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909864.402, "dur": 0.043,
    "args": {
      "Python parent id": 7790, "Python id": 7791, "finished": true, "Ev Idx": 10691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909864.590, "dur": 0.180,
    "args": {
      "Python parent id": 7790, "Python id": 7792, "finished": true, "Ev Idx": 10692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909865.796, "dur": 0.138,
    "args": {
      "Python parent id": 7783, "Python id": 7793, "finished": true, "Ev Idx": 10693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909866.213, "dur": 0.110,
    "args": {
      "Python parent id": 7781, "Python id": 7794, "finished": true, "Ev Idx": 10694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909866.608, "dur": 1.153,
    "args": {
      "Python parent id": 7781, "Python id": 7795, "finished": true, "Ev Idx": 10695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909866.788, "dur": 0.913,
    "args": {
      "Python parent id": 7795, "Python id": 7796, "finished": true, "Ev Idx": 10696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909866.937, "dur": 0.373,
    "args": {
      "Python parent id": 7796, "Python id": 7797, "finished": true, "Ev Idx": 10697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909867.080, "dur": 0.172,
    "args": {
      "Python parent id": 7797, "Python id": 7798, "finished": true, "Ev Idx": 10698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909867.168, "dur": 0.052,
    "args": {
      "Python parent id": 7798, "Python id": 7799, "finished": true, "Ev Idx": 10699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909867.428, "dur": 0.243,
    "args": {
      "Python parent id": 7796, "Python id": 7800, "finished": true, "Ev Idx": 10700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909868.112, "dur": 0.236,
    "args": {
      "Python parent id": 7765, "Python id": 7801, "finished": true, "Ev Idx": 10701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909868.950, "dur": 1.028,
    "args": {
      "Python parent id": 7765, "Python id": 7802, "finished": true, "Ev Idx": 10702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909869.414, "dur": 0.516,
    "args": {
      "Python parent id": 7802, "Python id": 7803, "finished": true, "Ev Idx": 10703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909870.420, "dur": 3.443,
    "args": {
      "Python parent id": 7764, "Python id": 7804, "finished": true, "Ev Idx": 10704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909874.465, "dur": 7.133,
    "args": {
      "Python parent id": 12, "Python id": 7805, "finished": true, "Ev Idx": 10705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740909882.021, "dur": 0.088,
    "args": {
      "Python parent id": 12, "Python id": 7806, "finished": true, "Ev Idx": 10706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909882.734, "dur": 0.173,
    "args": {
      "Python parent id": 12, "Python id": 7807, "finished": true, "Ev Idx": 10707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909883.378, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 7808, "finished": true, "Ev Idx": 10708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740909883.883, "dur": 9.585,
    "args": {
      "Python parent id": 12, "Python id": 7809, "finished": true, "Ev Idx": 10709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909894.834, "dur": 14.580,
    "args": {
      "Python parent id": 12, "Python id": 7810, "finished": true, "Ev Idx": 10710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909895.103, "dur": 10.694,
    "args": {
      "Python parent id": 7810, "Python id": 7811, "finished": true, "Ev Idx": 10711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909895.283, "dur": 0.415,
    "args": {
      "Python parent id": 7811, "Python id": 7812, "finished": true, "Ev Idx": 10712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909895.414, "dur": 0.218,
    "args": {
      "Python parent id": 7812, "Python id": 7813, "finished": true, "Ev Idx": 10713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909895.506, "dur": 0.090,
    "args": {
      "Python parent id": 7813, "Python id": 7814, "finished": true, "Ev Idx": 10714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909896.011, "dur": 8.027,
    "args": {
      "Python parent id": 7811, "Python id": 7815, "finished": true, "Ev Idx": 10715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909896.094, "dur": 0.065,
    "args": {
      "Python parent id": 7815, "Python id": 7816, "finished": true, "Ev Idx": 10716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909896.295, "dur": 0.055,
    "args": {
      "Python parent id": 7815, "Python id": 7817, "finished": true, "Ev Idx": 10717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909896.484, "dur": 0.060,
    "args": {
      "Python parent id": 7815, "Python id": 7818, "finished": true, "Ev Idx": 10718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909896.954, "dur": 0.058,
    "args": {
      "Python parent id": 7815, "Python id": 7819, "finished": true, "Ev Idx": 10719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909897.273, "dur": 0.071,
    "args": {
      "Python parent id": 7815, "Python id": 7820, "finished": true, "Ev Idx": 10720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909897.630, "dur": 6.329,
    "args": {
      "Python parent id": 7815, "Python id": 7821, "finished": true, "Ev Idx": 10721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909897.716, "dur": 0.057,
    "args": {
      "Python parent id": 7821, "Python id": 7822, "finished": true, "Ev Idx": 10722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909897.945, "dur": 0.062,
    "args": {
      "Python parent id": 7821, "Python id": 7823, "finished": true, "Ev Idx": 10723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909898.235, "dur": 0.064,
    "args": {
      "Python parent id": 7821, "Python id": 7824, "finished": true, "Ev Idx": 10724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740909898.628, "dur": 0.042,
    "args": {
      "Python parent id": 7821, "Python id": 7825, "finished": true, "Ev Idx": 10725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909898.990, "dur": 4.892,
    "args": {
      "Python parent id": 7821, "Python id": 7826, "finished": true, "Ev Idx": 10726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740909899.259, "dur": 4.550,
    "args": {
      "Python parent id": 7826, "Python id": 7827, "finished": true, "Ev Idx": 10727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740909899.420, "dur": 2.856,
    "args": {
      "Python parent id": 7827, "Python id": 7828, "finished": true, "Ev Idx": 10728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740909899.586, "dur": 2.631,
    "args": {
      "Python parent id": 7828, "Python id": 7829, "finished": true, "Ev Idx": 10729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740909899.712, "dur": 0.265,
    "args": {
      "Python parent id": 7829, "Python id": 7830, "finished": true, "Ev Idx": 10730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740909899.823, "dur": 0.067,
    "args": {
      "Python parent id": 7830, "Python id": 7831, "finished": true, "Ev Idx": 10731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.111, "dur": 1.792,
    "args": {
      "Python parent id": 7829, "Python id": 7832, "finished": true, "Ev Idx": 10732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.266, "dur": 1.584,
    "args": {
      "Python parent id": 7832, "Python id": 7833, "finished": true, "Ev Idx": 10733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.449, "dur": 1.356,
    "args": {
      "Python parent id": 7833, "Python id": 7834, "finished": true, "Ev Idx": 10734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.602, "dur": 1.019,
    "args": {
      "Python parent id": 7834, "Python id": 7835, "finished": true, "Ev Idx": 10735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.744, "dur": 0.444,
    "args": {
      "Python parent id": 7835, "Python id": 7836, "finished": true, "Ev Idx": 10736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909900.823, "dur": 0.070,
    "args": {
      "Python parent id": 7836, "Python id": 7837, "finished": true, "Ev Idx": 10737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740909901.037, "dur": 0.111,
    "args": {
      "Python parent id": 7836, "Python id": 7838, "finished": true, "Ev Idx": 10738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740909902.100, "dur": 0.069,
    "args": {
      "Python parent id": 7829, "Python id": 7839, "finished": true, "Ev Idx": 10739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740909902.419, "dur": 0.067,
    "args": {
      "Python parent id": 7827, "Python id": 7840, "finished": true, "Ev Idx": 10740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740909902.753, "dur": 1.017,
    "args": {
      "Python parent id": 7827, "Python id": 7841, "finished": true, "Ev Idx": 10741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740909902.903, "dur": 0.826,
    "args": {
      "Python parent id": 7841, "Python id": 7842, "finished": true, "Ev Idx": 10742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909903.026, "dur": 0.338,
    "args": {
      "Python parent id": 7842, "Python id": 7843, "finished": true, "Ev Idx": 10743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909903.142, "dur": 0.164,
    "args": {
      "Python parent id": 7843, "Python id": 7844, "finished": true, "Ev Idx": 10744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909903.228, "dur": 0.049,
    "args": {
      "Python parent id": 7844, "Python id": 7845, "finished": true, "Ev Idx": 10745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740909903.481, "dur": 0.219,
    "args": {
      "Python parent id": 7842, "Python id": 7846, "finished": true, "Ev Idx": 10746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740909904.162, "dur": 0.167,
    "args": {
      "Python parent id": 7811, "Python id": 7847, "finished": true, "Ev Idx": 10747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740909904.869, "dur": 0.812,
    "args": {
      "Python parent id": 7811, "Python id": 7848, "finished": true, "Ev Idx": 10748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740909905.230, "dur": 0.405,
    "args": {
      "Python parent id": 7848, "Python id": 7849, "finished": true, "Ev Idx": 10749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740909906.033, "dur": 3.325,
    "args": {
      "Python parent id": 7810, "Python id": 7850, "finished": true, "Ev Idx": 10750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740909910.263, "dur": 86.545,
    "args": {
      "Python parent id": 12, "Python id": 7851, "finished": true, "Ev Idx": 10751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909910.510, "dur": 0.426,
    "args": {
      "Python parent id": 7851, "Python id": 7852, "finished": true, "Ev Idx": 10752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909910.766, "dur": 0.079,
    "args": {
      "Python parent id": 7852, "Python id": 7853, "finished": true, "Ev Idx": 10753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909911.264, "dur": 0.294,
    "args": {
      "Python parent id": 7851, "Python id": 7854, "finished": true, "Ev Idx": 10754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740909911.740, "dur": 0.239,
    "args": {
      "Python parent id": 7851, "Python id": 7855, "finished": true, "Ev Idx": 10755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740909912.172, "dur": 1.585,
    "args": {
      "Python parent id": 7851, "Python id": 7856, "finished": true, "Ev Idx": 10756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740909914.323, "dur": 82.381,
    "args": {
      "Python parent id": 7851, "Python id": 7857, "finished": true, "Ev Idx": 10757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740909915.103, "dur": 81.423,
    "args": {
      "Python parent id": 7857, "Python id": 7858, "finished": true, "Ev Idx": 10758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909915.340, "dur": 81.132,
    "args": {
      "Python parent id": 7858, "Python id": 7859, "finished": true, "Ev Idx": 10759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740909920.103, "dur": 75.169,
    "args": {
      "Python parent id": 7859, "Python id": 7860, "finished": true, "Ev Idx": 10760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740909920.386, "dur": 74.806,
    "args": {
      "Python parent id": 7860, "Python id": 7861, "finished": true, "Ev Idx": 10761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740909920.897, "dur": 74.205,
    "args": {
      "Python parent id": 7861, "Python id": 7862, "finished": true, "Ev Idx": 10762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909921.410, "dur": 0.370,
    "args": {
      "Python parent id": 7862, "Python id": 7863, "finished": true, "Ev Idx": 10763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909921.677, "dur": 0.039,
    "args": {
      "Python parent id": 7863, "Python id": 7864, "finished": true, "Ev Idx": 10764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740909922.086, "dur": 0.300,
    "args": {
      "Python parent id": 7862, "Python id": 7865, "finished": true, "Ev Idx": 10765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740909922.301, "dur": 0.037,
    "args": {
      "Python parent id": 7865, "Python id": 7866, "finished": true, "Ev Idx": 10766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909922.676, "dur": 72.309,
    "args": {
      "Python parent id": 7862, "Python id": 7867, "finished": true, "Ev Idx": 10767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740909923.338, "dur": 70.406,
    "args": {
      "Python parent id": 7867, "Python id": 7868, "finished": true, "Ev Idx": 10768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740909929.917, "dur": 17.329,
    "args": {
      "Python parent id": 7868, "Python id": 7869, "finished": true, "Ev Idx": 10769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740909930.023, "dur": 0.124,
    "args": {
      "Python parent id": 7869, "Python id": 7870, "finished": true, "Ev Idx": 10770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740909930.370, "dur": 16.748,
    "args": {
      "Python parent id": 7869, "Python id": 7871, "finished": true, "Ev Idx": 10771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740909930.476, "dur": 7.102,
    "args": {
      "Python parent id": 7871, "Python id": 7872, "finished": true, "Ev Idx": 10772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909954.867, "dur": 7.051,
    "args": {
      "Python parent id": 7868, "Python id": 7873, "finished": true, "Ev Idx": 10773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740909970.898, "dur": 8.185,
    "args": {
      "Python parent id": 7868, "Python id": 7874, "finished": true, "Ev Idx": 10774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740909997.932, "dur": 17.118,
    "args": {
      "Python parent id": 12, "Python id": 7875, "finished": true, "Ev Idx": 10775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740909998.290, "dur": 11.310,
    "args": {
      "Python parent id": 7875, "Python id": 7876, "finished": true, "Ev Idx": 10776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740909998.504, "dur": 0.445,
    "args": {
      "Python parent id": 7876, "Python id": 7877, "finished": true, "Ev Idx": 10777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740909998.658, "dur": 0.226,
    "args": {
      "Python parent id": 7877, "Python id": 7878, "finished": true, "Ev Idx": 10778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740909998.779, "dur": 0.066,
    "args": {
      "Python parent id": 7878, "Python id": 7879, "finished": true, "Ev Idx": 10779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740909999.202, "dur": 8.410,
    "args": {
      "Python parent id": 7876, "Python id": 7880, "finished": true, "Ev Idx": 10780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909999.286, "dur": 0.069,
    "args": {
      "Python parent id": 7880, "Python id": 7881, "finished": true, "Ev Idx": 10781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909999.458, "dur": 0.059,
    "args": {
      "Python parent id": 7880, "Python id": 7882, "finished": true, "Ev Idx": 10782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740909999.682, "dur": 0.064,
    "args": {
      "Python parent id": 7880, "Python id": 7883, "finished": true, "Ev Idx": 10783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910000.220, "dur": 0.058,
    "args": {
      "Python parent id": 7880, "Python id": 7884, "finished": true, "Ev Idx": 10784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910000.588, "dur": 0.073,
    "args": {
      "Python parent id": 7880, "Python id": 7885, "finished": true, "Ev Idx": 10785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910001.059, "dur": 6.459,
    "args": {
      "Python parent id": 7880, "Python id": 7886, "finished": true, "Ev Idx": 10786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910001.188, "dur": 0.062,
    "args": {
      "Python parent id": 7886, "Python id": 7887, "finished": true, "Ev Idx": 10787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910001.391, "dur": 0.060,
    "args": {
      "Python parent id": 7886, "Python id": 7888, "finished": true, "Ev Idx": 10788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910001.605, "dur": 0.063,
    "args": {
      "Python parent id": 7886, "Python id": 7889, "finished": true, "Ev Idx": 10789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910001.954, "dur": 0.035,
    "args": {
      "Python parent id": 7886, "Python id": 7890, "finished": true, "Ev Idx": 10790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910002.254, "dur": 5.178,
    "args": {
      "Python parent id": 7886, "Python id": 7891, "finished": true, "Ev Idx": 10791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910002.542, "dur": 4.811,
    "args": {
      "Python parent id": 7891, "Python id": 7892, "finished": true, "Ev Idx": 10792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910002.664, "dur": 2.965,
    "args": {
      "Python parent id": 7892, "Python id": 7893, "finished": true, "Ev Idx": 10793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910002.844, "dur": 2.741,
    "args": {
      "Python parent id": 7893, "Python id": 7894, "finished": true, "Ev Idx": 10794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910002.969, "dur": 0.243,
    "args": {
      "Python parent id": 7894, "Python id": 7895, "finished": true, "Ev Idx": 10795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910003.075, "dur": 0.064,
    "args": {
      "Python parent id": 7895, "Python id": 7896, "finished": true, "Ev Idx": 10796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910003.357, "dur": 1.864,
    "args": {
      "Python parent id": 7894, "Python id": 7897, "finished": true, "Ev Idx": 10797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910003.550, "dur": 1.597,
    "args": {
      "Python parent id": 7897, "Python id": 7898, "finished": true, "Ev Idx": 10798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910003.753, "dur": 1.346,
    "args": {
      "Python parent id": 7898, "Python id": 7899, "finished": true, "Ev Idx": 10799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910003.872, "dur": 1.063,
    "args": {
      "Python parent id": 7899, "Python id": 7900, "finished": true, "Ev Idx": 10800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910004.041, "dur": 0.463,
    "args": {
      "Python parent id": 7900, "Python id": 7901, "finished": true, "Ev Idx": 10801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910004.123, "dur": 0.061,
    "args": {
      "Python parent id": 7901, "Python id": 7902, "finished": true, "Ev Idx": 10802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910004.332, "dur": 0.135,
    "args": {
      "Python parent id": 7901, "Python id": 7903, "finished": true, "Ev Idx": 10803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910005.419, "dur": 0.124,
    "args": {
      "Python parent id": 7894, "Python id": 7904, "finished": true, "Ev Idx": 10804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910005.804, "dur": 0.107,
    "args": {
      "Python parent id": 7892, "Python id": 7905, "finished": true, "Ev Idx": 10805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.214, "dur": 1.094,
    "args": {
      "Python parent id": 7892, "Python id": 7906, "finished": true, "Ev Idx": 10806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.358, "dur": 0.906,
    "args": {
      "Python parent id": 7906, "Python id": 7907, "finished": true, "Ev Idx": 10807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.504, "dur": 0.368,
    "args": {
      "Python parent id": 7907, "Python id": 7908, "finished": true, "Ev Idx": 10808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.643, "dur": 0.170,
    "args": {
      "Python parent id": 7908, "Python id": 7909, "finished": true, "Ev Idx": 10809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.725, "dur": 0.057,
    "args": {
      "Python parent id": 7909, "Python id": 7910, "finished": true, "Ev Idx": 10810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910006.998, "dur": 0.237,
    "args": {
      "Python parent id": 7907, "Python id": 7911, "finished": true, "Ev Idx": 10811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910007.730, "dur": 0.202,
    "args": {
      "Python parent id": 7876, "Python id": 7912, "finished": true, "Ev Idx": 10812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910008.533, "dur": 0.941,
    "args": {
      "Python parent id": 7876, "Python id": 7913, "finished": true, "Ev Idx": 10813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910008.947, "dur": 0.480,
    "args": {
      "Python parent id": 7913, "Python id": 7914, "finished": true, "Ev Idx": 10814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910009.866, "dur": 5.134,
    "args": {
      "Python parent id": 7875, "Python id": 7915, "finished": true, "Ev Idx": 10815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910015.611, "dur": 7.307,
    "args": {
      "Python parent id": 12, "Python id": 7916, "finished": true, "Ev Idx": 10816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910023.354, "dur": 0.111,
    "args": {
      "Python parent id": 12, "Python id": 7917, "finished": true, "Ev Idx": 10817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910024.125, "dur": 0.154,
    "args": {
      "Python parent id": 12, "Python id": 7918, "finished": true, "Ev Idx": 10818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910024.741, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 7919, "finished": true, "Ev Idx": 10819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910025.170, "dur": 11.232,
    "args": {
      "Python parent id": 12, "Python id": 7920, "finished": true, "Ev Idx": 10820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910037.754, "dur": 14.237,
    "args": {
      "Python parent id": 12, "Python id": 7921, "finished": true, "Ev Idx": 10821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910038.071, "dur": 10.366,
    "args": {
      "Python parent id": 7921, "Python id": 7922, "finished": true, "Ev Idx": 10822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910038.248, "dur": 0.459,
    "args": {
      "Python parent id": 7922, "Python id": 7923, "finished": true, "Ev Idx": 10823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910038.403, "dur": 0.242,
    "args": {
      "Python parent id": 7923, "Python id": 7924, "finished": true, "Ev Idx": 10824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910038.502, "dur": 0.090,
    "args": {
      "Python parent id": 7924, "Python id": 7925, "finished": true, "Ev Idx": 10825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910038.960, "dur": 7.834,
    "args": {
      "Python parent id": 7922, "Python id": 7926, "finished": true, "Ev Idx": 10826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910039.044, "dur": 0.106,
    "args": {
      "Python parent id": 7926, "Python id": 7927, "finished": true, "Ev Idx": 10827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910039.241, "dur": 0.060,
    "args": {
      "Python parent id": 7926, "Python id": 7928, "finished": true, "Ev Idx": 10828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910039.462, "dur": 0.059,
    "args": {
      "Python parent id": 7926, "Python id": 7929, "finished": true, "Ev Idx": 10829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910039.967, "dur": 0.040,
    "args": {
      "Python parent id": 7926, "Python id": 7930, "finished": true, "Ev Idx": 10830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910040.253, "dur": 0.069,
    "args": {
      "Python parent id": 7926, "Python id": 7931, "finished": true, "Ev Idx": 10831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910040.688, "dur": 6.012,
    "args": {
      "Python parent id": 7926, "Python id": 7932, "finished": true, "Ev Idx": 10832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910040.785, "dur": 0.066,
    "args": {
      "Python parent id": 7932, "Python id": 7933, "finished": true, "Ev Idx": 10833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910040.990, "dur": 0.051,
    "args": {
      "Python parent id": 7932, "Python id": 7934, "finished": true, "Ev Idx": 10834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910041.250, "dur": 0.062,
    "args": {
      "Python parent id": 7932, "Python id": 7935, "finished": true, "Ev Idx": 10835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910041.675, "dur": 0.037,
    "args": {
      "Python parent id": 7932, "Python id": 7936, "finished": true, "Ev Idx": 10836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.014, "dur": 4.632,
    "args": {
      "Python parent id": 7932, "Python id": 7937, "finished": true, "Ev Idx": 10837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.242, "dur": 4.351,
    "args": {
      "Python parent id": 7937, "Python id": 7938, "finished": true, "Ev Idx": 10838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.361, "dur": 2.693,
    "args": {
      "Python parent id": 7938, "Python id": 7939, "finished": true, "Ev Idx": 10839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.522, "dur": 2.490,
    "args": {
      "Python parent id": 7939, "Python id": 7940, "finished": true, "Ev Idx": 10840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.643, "dur": 0.223,
    "args": {
      "Python parent id": 7940, "Python id": 7941, "finished": true, "Ev Idx": 10841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.748, "dur": 0.068,
    "args": {
      "Python parent id": 7941, "Python id": 7942, "finished": true, "Ev Idx": 10842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910042.996, "dur": 1.717,
    "args": {
      "Python parent id": 7940, "Python id": 7943, "finished": true, "Ev Idx": 10843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.149, "dur": 1.510,
    "args": {
      "Python parent id": 7943, "Python id": 7944, "finished": true, "Ev Idx": 10844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.300, "dur": 1.319,
    "args": {
      "Python parent id": 7944, "Python id": 7945, "finished": true, "Ev Idx": 10845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.475, "dur": 0.959,
    "args": {
      "Python parent id": 7945, "Python id": 7946, "finished": true, "Ev Idx": 10846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.613, "dur": 0.428,
    "args": {
      "Python parent id": 7946, "Python id": 7947, "finished": true, "Ev Idx": 10847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.686, "dur": 0.065,
    "args": {
      "Python parent id": 7947, "Python id": 7948, "finished": true, "Ev Idx": 10848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910043.875, "dur": 0.116,
    "args": {
      "Python parent id": 7947, "Python id": 7949, "finished": true, "Ev Idx": 10849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910044.882, "dur": 0.100,
    "args": {
      "Python parent id": 7940, "Python id": 7950, "finished": true, "Ev Idx": 10850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910045.196, "dur": 0.092,
    "args": {
      "Python parent id": 7938, "Python id": 7951, "finished": true, "Ev Idx": 10851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910045.541, "dur": 1.009,
    "args": {
      "Python parent id": 7938, "Python id": 7952, "finished": true, "Ev Idx": 10852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910045.660, "dur": 0.849,
    "args": {
      "Python parent id": 7952, "Python id": 7953, "finished": true, "Ev Idx": 10853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910045.788, "dur": 0.364,
    "args": {
      "Python parent id": 7953, "Python id": 7954, "finished": true, "Ev Idx": 10854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910045.920, "dur": 0.174,
    "args": {
      "Python parent id": 7954, "Python id": 7955, "finished": true, "Ev Idx": 10855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910046.000, "dur": 0.062,
    "args": {
      "Python parent id": 7955, "Python id": 7956, "finished": true, "Ev Idx": 10856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910046.269, "dur": 0.209,
    "args": {
      "Python parent id": 7953, "Python id": 7957, "finished": true, "Ev Idx": 10857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910046.898, "dur": 0.124,
    "args": {
      "Python parent id": 7922, "Python id": 7958, "finished": true, "Ev Idx": 10858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910047.496, "dur": 0.796,
    "args": {
      "Python parent id": 7922, "Python id": 7959, "finished": true, "Ev Idx": 10859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910047.876, "dur": 0.367,
    "args": {
      "Python parent id": 7959, "Python id": 7960, "finished": true, "Ev Idx": 10860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910048.655, "dur": 3.278,
    "args": {
      "Python parent id": 7921, "Python id": 7961, "finished": true, "Ev Idx": 10861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910052.800, "dur": 131.548,
    "args": {
      "Python parent id": 12, "Python id": 7962, "finished": true, "Ev Idx": 10862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910053.068, "dur": 0.382,
    "args": {
      "Python parent id": 7962, "Python id": 7963, "finished": true, "Ev Idx": 10863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910053.312, "dur": 0.077,
    "args": {
      "Python parent id": 7963, "Python id": 7964, "finished": true, "Ev Idx": 10864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910053.724, "dur": 0.296,
    "args": {
      "Python parent id": 7962, "Python id": 7965, "finished": true, "Ev Idx": 10865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910054.191, "dur": 0.288,
    "args": {
      "Python parent id": 7962, "Python id": 7966, "finished": true, "Ev Idx": 10866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910054.684, "dur": 1.435,
    "args": {
      "Python parent id": 7962, "Python id": 7967, "finished": true, "Ev Idx": 10867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910056.690, "dur": 127.552,
    "args": {
      "Python parent id": 7962, "Python id": 7968, "finished": true, "Ev Idx": 10868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910057.574, "dur": 126.424,
    "args": {
      "Python parent id": 7968, "Python id": 7969, "finished": true, "Ev Idx": 10869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910057.828, "dur": 126.073,
    "args": {
      "Python parent id": 7969, "Python id": 7970, "finished": true, "Ev Idx": 10870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910063.401, "dur": 118.734,
    "args": {
      "Python parent id": 7970, "Python id": 7971, "finished": true, "Ev Idx": 10871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910063.656, "dur": 118.427,
    "args": {
      "Python parent id": 7971, "Python id": 7972, "finished": true, "Ev Idx": 10872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910064.130, "dur": 117.834,
    "args": {
      "Python parent id": 7972, "Python id": 7973, "finished": true, "Ev Idx": 10873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910064.573, "dur": 0.301,
    "args": {
      "Python parent id": 7973, "Python id": 7974, "finished": true, "Ev Idx": 10874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910064.784, "dur": 0.038,
    "args": {
      "Python parent id": 7974, "Python id": 7975, "finished": true, "Ev Idx": 10875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910065.183, "dur": 0.386,
    "args": {
      "Python parent id": 7973, "Python id": 7976, "finished": true, "Ev Idx": 10876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910065.461, "dur": 0.069,
    "args": {
      "Python parent id": 7976, "Python id": 7977, "finished": true, "Ev Idx": 10877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910065.880, "dur": 115.881,
    "args": {
      "Python parent id": 7973, "Python id": 7978, "finished": true, "Ev Idx": 10878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910066.505, "dur": 113.780,
    "args": {
      "Python parent id": 7978, "Python id": 7979, "finished": true, "Ev Idx": 10879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910090.477, "dur": 24.554,
    "args": {
      "Python parent id": 7979, "Python id": 7980, "finished": true, "Ev Idx": 10880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910090.675, "dur": 0.145,
    "args": {
      "Python parent id": 7980, "Python id": 7981, "finished": true, "Ev Idx": 10881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910091.225, "dur": 23.572,
    "args": {
      "Python parent id": 7980, "Python id": 7982, "finished": true, "Ev Idx": 10882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910091.388, "dur": 11.597,
    "args": {
      "Python parent id": 7982, "Python id": 7983, "finished": true, "Ev Idx": 10883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910122.913, "dur": 7.180,
    "args": {
      "Python parent id": 7979, "Python id": 7984, "finished": true, "Ev Idx": 10884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910143.267, "dur": 10.327,
    "args": {
      "Python parent id": 7979, "Python id": 7985, "finished": true, "Ev Idx": 10885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910185.793, "dur": 19.408,
    "args": {
      "Python parent id": 12, "Python id": 7986, "finished": true, "Ev Idx": 10886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910186.144, "dur": 14.645,
    "args": {
      "Python parent id": 7986, "Python id": 7987, "finished": true, "Ev Idx": 10887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910186.378, "dur": 0.680,
    "args": {
      "Python parent id": 7987, "Python id": 7988, "finished": true, "Ev Idx": 10888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910186.529, "dur": 0.423,
    "args": {
      "Python parent id": 7988, "Python id": 7989, "finished": true, "Ev Idx": 10889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910186.735, "dur": 0.160,
    "args": {
      "Python parent id": 7989, "Python id": 7990, "finished": true, "Ev Idx": 10890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910187.456, "dur": 10.477,
    "args": {
      "Python parent id": 7987, "Python id": 7991, "finished": true, "Ev Idx": 10891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910187.582, "dur": 0.155,
    "args": {
      "Python parent id": 7991, "Python id": 7992, "finished": true, "Ev Idx": 10892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910187.854, "dur": 0.058,
    "args": {
      "Python parent id": 7991, "Python id": 7993, "finished": true, "Ev Idx": 10893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910188.054, "dur": 0.056,
    "args": {
      "Python parent id": 7991, "Python id": 7994, "finished": true, "Ev Idx": 10894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910188.583, "dur": 0.068,
    "args": {
      "Python parent id": 7991, "Python id": 7995, "finished": true, "Ev Idx": 10895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910188.927, "dur": 0.083,
    "args": {
      "Python parent id": 7991, "Python id": 7996, "finished": true, "Ev Idx": 10896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910189.362, "dur": 8.525,
    "args": {
      "Python parent id": 7991, "Python id": 7997, "finished": true, "Ev Idx": 10897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910189.455, "dur": 0.060,
    "args": {
      "Python parent id": 7997, "Python id": 7998, "finished": true, "Ev Idx": 10898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910189.688, "dur": 0.057,
    "args": {
      "Python parent id": 7997, "Python id": 7999, "finished": true, "Ev Idx": 10899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910189.922, "dur": 0.079,
    "args": {
      "Python parent id": 7997, "Python id": 8000, "finished": true, "Ev Idx": 10900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910190.373, "dur": 0.037,
    "args": {
      "Python parent id": 7997, "Python id": 8001, "finished": true, "Ev Idx": 10901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910190.790, "dur": 6.986,
    "args": {
      "Python parent id": 7997, "Python id": 8002, "finished": true, "Ev Idx": 10902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.036, "dur": 6.642,
    "args": {
      "Python parent id": 8002, "Python id": 8003, "finished": true, "Ev Idx": 10903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.230, "dur": 4.439,
    "args": {
      "Python parent id": 8003, "Python id": 8004, "finished": true, "Ev Idx": 10904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.402, "dur": 4.197,
    "args": {
      "Python parent id": 8004, "Python id": 8005, "finished": true, "Ev Idx": 10905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.533, "dur": 0.311,
    "args": {
      "Python parent id": 8005, "Python id": 8006, "finished": true, "Ev Idx": 10906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.675, "dur": 0.085,
    "args": {
      "Python parent id": 8006, "Python id": 8007, "finished": true, "Ev Idx": 10907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910191.999, "dur": 3.169,
    "args": {
      "Python parent id": 8005, "Python id": 8008, "finished": true, "Ev Idx": 10908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910192.254, "dur": 2.784,
    "args": {
      "Python parent id": 8008, "Python id": 8009, "finished": true, "Ev Idx": 10909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910192.530, "dur": 2.430,
    "args": {
      "Python parent id": 8009, "Python id": 8010, "finished": true, "Ev Idx": 10910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910192.783, "dur": 1.809,
    "args": {
      "Python parent id": 8010, "Python id": 8011, "finished": true, "Ev Idx": 10911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910193.032, "dur": 0.740,
    "args": {
      "Python parent id": 8011, "Python id": 8012, "finished": true, "Ev Idx": 10912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910193.141, "dur": 0.058,
    "args": {
      "Python parent id": 8012, "Python id": 8013, "finished": true, "Ev Idx": 10913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910193.423, "dur": 0.288,
    "args": {
      "Python parent id": 8012, "Python id": 8014, "finished": true, "Ev Idx": 10914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910195.352, "dur": 0.184,
    "args": {
      "Python parent id": 8005, "Python id": 8015, "finished": true, "Ev Idx": 10915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910195.905, "dur": 0.130,
    "args": {
      "Python parent id": 8003, "Python id": 8016, "finished": true, "Ev Idx": 10916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910196.363, "dur": 1.258,
    "args": {
      "Python parent id": 8003, "Python id": 8017, "finished": true, "Ev Idx": 10917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910196.499, "dur": 1.052,
    "args": {
      "Python parent id": 8017, "Python id": 8018, "finished": true, "Ev Idx": 10918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910196.614, "dur": 0.400,
    "args": {
      "Python parent id": 8018, "Python id": 8019, "finished": true, "Ev Idx": 10919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910196.739, "dur": 0.204,
    "args": {
      "Python parent id": 8019, "Python id": 8020, "finished": true, "Ev Idx": 10920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910196.845, "dur": 0.069,
    "args": {
      "Python parent id": 8020, "Python id": 8021, "finished": true, "Ev Idx": 10921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910197.163, "dur": 0.332,
    "args": {
      "Python parent id": 8018, "Python id": 8022, "finished": true, "Ev Idx": 10922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910198.048, "dur": 0.297,
    "args": {
      "Python parent id": 7987, "Python id": 8023, "finished": true, "Ev Idx": 10923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910199.161, "dur": 1.462,
    "args": {
      "Python parent id": 7987, "Python id": 8024, "finished": true, "Ev Idx": 10924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910199.854, "dur": 0.687,
    "args": {
      "Python parent id": 8024, "Python id": 8025, "finished": true, "Ev Idx": 10925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910201.186, "dur": 3.897,
    "args": {
      "Python parent id": 7986, "Python id": 8026, "finished": true, "Ev Idx": 10926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910205.804, "dur": 8.801,
    "args": {
      "Python parent id": 12, "Python id": 8027, "finished": true, "Ev Idx": 10927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910215.046, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 8028, "finished": true, "Ev Idx": 10928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910216.038, "dur": 0.225,
    "args": {
      "Python parent id": 12, "Python id": 8029, "finished": true, "Ev Idx": 10929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910216.831, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 8030, "finished": true, "Ev Idx": 10930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910217.315, "dur": 10.922,
    "args": {
      "Python parent id": 12, "Python id": 8031, "finished": true, "Ev Idx": 10931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910229.496, "dur": 15.758,
    "args": {
      "Python parent id": 12, "Python id": 8032, "finished": true, "Ev Idx": 10932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910229.787, "dur": 11.535,
    "args": {
      "Python parent id": 8032, "Python id": 8033, "finished": true, "Ev Idx": 10933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910229.928, "dur": 0.509,
    "args": {
      "Python parent id": 8033, "Python id": 8034, "finished": true, "Ev Idx": 10934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910230.080, "dur": 0.272,
    "args": {
      "Python parent id": 8034, "Python id": 8035, "finished": true, "Ev Idx": 10935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910230.210, "dur": 0.111,
    "args": {
      "Python parent id": 8035, "Python id": 8036, "finished": true, "Ev Idx": 10936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910230.769, "dur": 8.577,
    "args": {
      "Python parent id": 8033, "Python id": 8037, "finished": true, "Ev Idx": 10937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910230.883, "dur": 0.063,
    "args": {
      "Python parent id": 8037, "Python id": 8038, "finished": true, "Ev Idx": 10938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910231.048, "dur": 0.080,
    "args": {
      "Python parent id": 8037, "Python id": 8039, "finished": true, "Ev Idx": 10939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910231.259, "dur": 0.057,
    "args": {
      "Python parent id": 8037, "Python id": 8040, "finished": true, "Ev Idx": 10940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910231.827, "dur": 0.039,
    "args": {
      "Python parent id": 8037, "Python id": 8041, "finished": true, "Ev Idx": 10941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910232.133, "dur": 0.063,
    "args": {
      "Python parent id": 8037, "Python id": 8042, "finished": true, "Ev Idx": 10942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910232.552, "dur": 6.743,
    "args": {
      "Python parent id": 8037, "Python id": 8043, "finished": true, "Ev Idx": 10943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910232.622, "dur": 0.074,
    "args": {
      "Python parent id": 8043, "Python id": 8044, "finished": true, "Ev Idx": 10944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910232.851, "dur": 0.055,
    "args": {
      "Python parent id": 8043, "Python id": 8045, "finished": true, "Ev Idx": 10945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910233.056, "dur": 0.059,
    "args": {
      "Python parent id": 8043, "Python id": 8046, "finished": true, "Ev Idx": 10946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910233.480, "dur": 0.029,
    "args": {
      "Python parent id": 8043, "Python id": 8047, "finished": true, "Ev Idx": 10947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910233.776, "dur": 5.427,
    "args": {
      "Python parent id": 8043, "Python id": 8048, "finished": true, "Ev Idx": 10948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.025, "dur": 5.081,
    "args": {
      "Python parent id": 8048, "Python id": 8049, "finished": true, "Ev Idx": 10949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.152, "dur": 3.197,
    "args": {
      "Python parent id": 8049, "Python id": 8050, "finished": true, "Ev Idx": 10950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.347, "dur": 2.936,
    "args": {
      "Python parent id": 8050, "Python id": 8051, "finished": true, "Ev Idx": 10951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.488, "dur": 0.251,
    "args": {
      "Python parent id": 8051, "Python id": 8052, "finished": true, "Ev Idx": 10952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.601, "dur": 0.057,
    "args": {
      "Python parent id": 8052, "Python id": 8053, "finished": true, "Ev Idx": 10953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910234.891, "dur": 2.056,
    "args": {
      "Python parent id": 8051, "Python id": 8054, "finished": true, "Ev Idx": 10954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.063, "dur": 1.803,
    "args": {
      "Python parent id": 8054, "Python id": 8055, "finished": true, "Ev Idx": 10955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.250, "dur": 1.554,
    "args": {
      "Python parent id": 8055, "Python id": 8056, "finished": true, "Ev Idx": 10956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.406, "dur": 1.184,
    "args": {
      "Python parent id": 8056, "Python id": 8057, "finished": true, "Ev Idx": 10957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.597, "dur": 0.519,
    "args": {
      "Python parent id": 8057, "Python id": 8058, "finished": true, "Ev Idx": 10958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.701, "dur": 0.031,
    "args": {
      "Python parent id": 8058, "Python id": 8059, "finished": true, "Ev Idx": 10959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910235.906, "dur": 0.168,
    "args": {
      "Python parent id": 8058, "Python id": 8060, "finished": true, "Ev Idx": 10960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910237.128, "dur": 0.098,
    "args": {
      "Python parent id": 8051, "Python id": 8061, "finished": true, "Ev Idx": 10961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910237.520, "dur": 0.089,
    "args": {
      "Python parent id": 8049, "Python id": 8062, "finished": true, "Ev Idx": 10962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910237.864, "dur": 1.204,
    "args": {
      "Python parent id": 8049, "Python id": 8063, "finished": true, "Ev Idx": 10963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910238.008, "dur": 1.008,
    "args": {
      "Python parent id": 8063, "Python id": 8064, "finished": true, "Ev Idx": 10964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910238.193, "dur": 0.366,
    "args": {
      "Python parent id": 8064, "Python id": 8065, "finished": true, "Ev Idx": 10965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910238.322, "dur": 0.177,
    "args": {
      "Python parent id": 8065, "Python id": 8066, "finished": true, "Ev Idx": 10966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910238.414, "dur": 0.054,
    "args": {
      "Python parent id": 8066, "Python id": 8067, "finished": true, "Ev Idx": 10967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910238.671, "dur": 0.297,
    "args": {
      "Python parent id": 8064, "Python id": 8068, "finished": true, "Ev Idx": 10968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910239.457, "dur": 0.157,
    "args": {
      "Python parent id": 8033, "Python id": 8069, "finished": true, "Ev Idx": 10969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910240.245, "dur": 0.930,
    "args": {
      "Python parent id": 8033, "Python id": 8070, "finished": true, "Ev Idx": 10970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910240.693, "dur": 0.436,
    "args": {
      "Python parent id": 8070, "Python id": 8071, "finished": true, "Ev Idx": 10971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910241.604, "dur": 3.602,
    "args": {
      "Python parent id": 8032, "Python id": 8072, "finished": true, "Ev Idx": 10972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910246.156, "dur": 92.570,
    "args": {
      "Python parent id": 12, "Python id": 8073, "finished": true, "Ev Idx": 10973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910246.603, "dur": 0.580,
    "args": {
      "Python parent id": 8073, "Python id": 8074, "finished": true, "Ev Idx": 10974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910246.962, "dur": 0.162,
    "args": {
      "Python parent id": 8074, "Python id": 8075, "finished": true, "Ev Idx": 10975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910247.467, "dur": 0.636,
    "args": {
      "Python parent id": 8073, "Python id": 8076, "finished": true, "Ev Idx": 10976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910248.357, "dur": 0.267,
    "args": {
      "Python parent id": 8073, "Python id": 8077, "finished": true, "Ev Idx": 10977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910248.849, "dur": 1.553,
    "args": {
      "Python parent id": 8073, "Python id": 8078, "finished": true, "Ev Idx": 10978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910251.024, "dur": 87.638,
    "args": {
      "Python parent id": 8073, "Python id": 8079, "finished": true, "Ev Idx": 10979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910252.001, "dur": 86.421,
    "args": {
      "Python parent id": 8079, "Python id": 8080, "finished": true, "Ev Idx": 10980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910252.351, "dur": 86.008,
    "args": {
      "Python parent id": 8080, "Python id": 8081, "finished": true, "Ev Idx": 10981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910258.506, "dur": 78.445,
    "args": {
      "Python parent id": 8081, "Python id": 8082, "finished": true, "Ev Idx": 10982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910258.817, "dur": 78.063,
    "args": {
      "Python parent id": 8082, "Python id": 8083, "finished": true, "Ev Idx": 10983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910259.412, "dur": 77.259,
    "args": {
      "Python parent id": 8083, "Python id": 8084, "finished": true, "Ev Idx": 10984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910259.979, "dur": 0.415,
    "args": {
      "Python parent id": 8084, "Python id": 8085, "finished": true, "Ev Idx": 10985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910260.267, "dur": 0.085,
    "args": {
      "Python parent id": 8085, "Python id": 8086, "finished": true, "Ev Idx": 10986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910260.696, "dur": 0.343,
    "args": {
      "Python parent id": 8084, "Python id": 8087, "finished": true, "Ev Idx": 10987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910260.944, "dur": 0.045,
    "args": {
      "Python parent id": 8087, "Python id": 8088, "finished": true, "Ev Idx": 10988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910261.365, "dur": 75.218,
    "args": {
      "Python parent id": 8084, "Python id": 8089, "finished": true, "Ev Idx": 10989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910262.142, "dur": 73.146,
    "args": {
      "Python parent id": 8089, "Python id": 8090, "finished": true, "Ev Idx": 10990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910269.594, "dur": 18.327,
    "args": {
      "Python parent id": 8090, "Python id": 8091, "finished": true, "Ev Idx": 10991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910269.752, "dur": 0.134,
    "args": {
      "Python parent id": 8091, "Python id": 8092, "finished": true, "Ev Idx": 10992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910270.198, "dur": 17.546,
    "args": {
      "Python parent id": 8091, "Python id": 8093, "finished": true, "Ev Idx": 10993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910270.324, "dur": 7.789,
    "args": {
      "Python parent id": 8093, "Python id": 8094, "finished": true, "Ev Idx": 10994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910295.947, "dur": 7.477,
    "args": {
      "Python parent id": 8090, "Python id": 8095, "finished": true, "Ev Idx": 10995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910312.367, "dur": 8.313,
    "args": {
      "Python parent id": 8090, "Python id": 8096, "finished": true, "Ev Idx": 10996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910339.912, "dur": 17.778,
    "args": {
      "Python parent id": 12, "Python id": 8097, "finished": true, "Ev Idx": 10997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910340.280, "dur": 12.938,
    "args": {
      "Python parent id": 8097, "Python id": 8098, "finished": true, "Ev Idx": 10998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910340.494, "dur": 0.500,
    "args": {
      "Python parent id": 8098, "Python id": 8099, "finished": true, "Ev Idx": 10999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910340.654, "dur": 0.269,
    "args": {
      "Python parent id": 8099, "Python id": 8100, "finished": true, "Ev Idx": 11000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910340.772, "dur": 0.113,
    "args": {
      "Python parent id": 8100, "Python id": 8101, "finished": true, "Ev Idx": 11001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910341.343, "dur": 9.531,
    "args": {
      "Python parent id": 8098, "Python id": 8102, "finished": true, "Ev Idx": 11002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910341.424, "dur": 0.065,
    "args": {
      "Python parent id": 8102, "Python id": 8103, "finished": true, "Ev Idx": 11003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910341.645, "dur": 0.059,
    "args": {
      "Python parent id": 8102, "Python id": 8104, "finished": true, "Ev Idx": 11004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910341.855, "dur": 0.062,
    "args": {
      "Python parent id": 8102, "Python id": 8105, "finished": true, "Ev Idx": 11005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910342.376, "dur": 0.058,
    "args": {
      "Python parent id": 8102, "Python id": 8106, "finished": true, "Ev Idx": 11006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910342.723, "dur": 0.069,
    "args": {
      "Python parent id": 8102, "Python id": 8107, "finished": true, "Ev Idx": 11007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910343.177, "dur": 7.617,
    "args": {
      "Python parent id": 8102, "Python id": 8108, "finished": true, "Ev Idx": 11008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910343.255, "dur": 0.060,
    "args": {
      "Python parent id": 8108, "Python id": 8109, "finished": true, "Ev Idx": 11009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910343.495, "dur": 0.062,
    "args": {
      "Python parent id": 8108, "Python id": 8110, "finished": true, "Ev Idx": 11010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910343.724, "dur": 0.069,
    "args": {
      "Python parent id": 8108, "Python id": 8111, "finished": true, "Ev Idx": 11011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910344.154, "dur": 0.055,
    "args": {
      "Python parent id": 8108, "Python id": 8112, "finished": true, "Ev Idx": 11012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910344.539, "dur": 6.154,
    "args": {
      "Python parent id": 8108, "Python id": 8113, "finished": true, "Ev Idx": 11013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910344.767, "dur": 5.871,
    "args": {
      "Python parent id": 8113, "Python id": 8114, "finished": true, "Ev Idx": 11014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910344.912, "dur": 3.746,
    "args": {
      "Python parent id": 8114, "Python id": 8115, "finished": true, "Ev Idx": 11015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910345.105, "dur": 3.473,
    "args": {
      "Python parent id": 8115, "Python id": 8116, "finished": true, "Ev Idx": 11016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910345.279, "dur": 0.266,
    "args": {
      "Python parent id": 8116, "Python id": 8117, "finished": true, "Ev Idx": 11017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910345.408, "dur": 0.064,
    "args": {
      "Python parent id": 8117, "Python id": 8118, "finished": true, "Ev Idx": 11018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910345.712, "dur": 2.476,
    "args": {
      "Python parent id": 8116, "Python id": 8119, "finished": true, "Ev Idx": 11019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910345.926, "dur": 2.154,
    "args": {
      "Python parent id": 8119, "Python id": 8120, "finished": true, "Ev Idx": 11020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910346.125, "dur": 1.892,
    "args": {
      "Python parent id": 8120, "Python id": 8121, "finished": true, "Ev Idx": 11021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910346.282, "dur": 1.464,
    "args": {
      "Python parent id": 8121, "Python id": 8122, "finished": true, "Ev Idx": 11022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910346.524, "dur": 0.624,
    "args": {
      "Python parent id": 8122, "Python id": 8123, "finished": true, "Ev Idx": 11023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910346.613, "dur": 0.060,
    "args": {
      "Python parent id": 8123, "Python id": 8124, "finished": true, "Ev Idx": 11024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910346.912, "dur": 0.177,
    "args": {
      "Python parent id": 8123, "Python id": 8125, "finished": true, "Ev Idx": 11025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910348.363, "dur": 0.159,
    "args": {
      "Python parent id": 8116, "Python id": 8126, "finished": true, "Ev Idx": 11026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910348.853, "dur": 0.116,
    "args": {
      "Python parent id": 8114, "Python id": 8127, "finished": true, "Ev Idx": 11027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910349.290, "dur": 1.312,
    "args": {
      "Python parent id": 8114, "Python id": 8128, "finished": true, "Ev Idx": 11028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910349.469, "dur": 1.059,
    "args": {
      "Python parent id": 8128, "Python id": 8129, "finished": true, "Ev Idx": 11029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910349.595, "dur": 0.379,
    "args": {
      "Python parent id": 8129, "Python id": 8130, "finished": true, "Ev Idx": 11030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910349.727, "dur": 0.192,
    "args": {
      "Python parent id": 8130, "Python id": 8131, "finished": true, "Ev Idx": 11031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910349.816, "dur": 0.071,
    "args": {
      "Python parent id": 8131, "Python id": 8132, "finished": true, "Ev Idx": 11032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910350.197, "dur": 0.274,
    "args": {
      "Python parent id": 8129, "Python id": 8133, "finished": true, "Ev Idx": 11033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910351.030, "dur": 0.292,
    "args": {
      "Python parent id": 8098, "Python id": 8134, "finished": true, "Ev Idx": 11034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910351.949, "dur": 1.120,
    "args": {
      "Python parent id": 8098, "Python id": 8135, "finished": true, "Ev Idx": 11035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910352.429, "dur": 0.570,
    "args": {
      "Python parent id": 8135, "Python id": 8136, "finished": true, "Ev Idx": 11036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910353.511, "dur": 4.079,
    "args": {
      "Python parent id": 8097, "Python id": 8137, "finished": true, "Ev Idx": 11037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910358.317, "dur": 7.436,
    "args": {
      "Python parent id": 12, "Python id": 8138, "finished": true, "Ev Idx": 11038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910366.202, "dur": 0.108,
    "args": {
      "Python parent id": 12, "Python id": 8139, "finished": true, "Ev Idx": 11039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910366.929, "dur": 0.165,
    "args": {
      "Python parent id": 12, "Python id": 8140, "finished": true, "Ev Idx": 11040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910367.569, "dur": 0.066,
    "args": {
      "Python parent id": 12, "Python id": 8141, "finished": true, "Ev Idx": 11041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910368.001, "dur": 10.133,
    "args": {
      "Python parent id": 12, "Python id": 8142, "finished": true, "Ev Idx": 11042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910379.566, "dur": 15.101,
    "args": {
      "Python parent id": 12, "Python id": 8143, "finished": true, "Ev Idx": 11043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910379.829, "dur": 11.166,
    "args": {
      "Python parent id": 8143, "Python id": 8144, "finished": true, "Ev Idx": 11044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910379.969, "dur": 0.505,
    "args": {
      "Python parent id": 8144, "Python id": 8145, "finished": true, "Ev Idx": 11045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910380.116, "dur": 0.278,
    "args": {
      "Python parent id": 8145, "Python id": 8146, "finished": true, "Ev Idx": 11046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910380.236, "dur": 0.103,
    "args": {
      "Python parent id": 8146, "Python id": 8147, "finished": true, "Ev Idx": 11047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910380.764, "dur": 8.478,
    "args": {
      "Python parent id": 8144, "Python id": 8148, "finished": true, "Ev Idx": 11048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910380.854, "dur": 0.065,
    "args": {
      "Python parent id": 8148, "Python id": 8149, "finished": true, "Ev Idx": 11049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910381.046, "dur": 0.052,
    "args": {
      "Python parent id": 8148, "Python id": 8150, "finished": true, "Ev Idx": 11050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910381.231, "dur": 0.075,
    "args": {
      "Python parent id": 8148, "Python id": 8151, "finished": true, "Ev Idx": 11051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910381.711, "dur": 0.046,
    "args": {
      "Python parent id": 8148, "Python id": 8152, "finished": true, "Ev Idx": 11052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910382.048, "dur": 0.064,
    "args": {
      "Python parent id": 8148, "Python id": 8153, "finished": true, "Ev Idx": 11053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910382.482, "dur": 6.681,
    "args": {
      "Python parent id": 8148, "Python id": 8154, "finished": true, "Ev Idx": 11054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910382.558, "dur": 0.054,
    "args": {
      "Python parent id": 8154, "Python id": 8155, "finished": true, "Ev Idx": 11055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910382.782, "dur": 0.048,
    "args": {
      "Python parent id": 8154, "Python id": 8156, "finished": true, "Ev Idx": 11056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910383.007, "dur": 0.062,
    "args": {
      "Python parent id": 8154, "Python id": 8157, "finished": true, "Ev Idx": 11057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910383.455, "dur": 0.025,
    "args": {
      "Python parent id": 8154, "Python id": 8158, "finished": true, "Ev Idx": 11058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910383.851, "dur": 5.258,
    "args": {
      "Python parent id": 8154, "Python id": 8159, "finished": true, "Ev Idx": 11059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.124, "dur": 4.933,
    "args": {
      "Python parent id": 8159, "Python id": 8160, "finished": true, "Ev Idx": 11060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.255, "dur": 3.099,
    "args": {
      "Python parent id": 8160, "Python id": 8161, "finished": true, "Ev Idx": 11061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.435, "dur": 2.878,
    "args": {
      "Python parent id": 8161, "Python id": 8162, "finished": true, "Ev Idx": 11062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.570, "dur": 0.270,
    "args": {
      "Python parent id": 8162, "Python id": 8163, "finished": true, "Ev Idx": 11063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.682, "dur": 0.067,
    "args": {
      "Python parent id": 8163, "Python id": 8164, "finished": true, "Ev Idx": 11064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910384.973, "dur": 1.991,
    "args": {
      "Python parent id": 8162, "Python id": 8165, "finished": true, "Ev Idx": 11065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.179, "dur": 1.704,
    "args": {
      "Python parent id": 8165, "Python id": 8166, "finished": true, "Ev Idx": 11066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.330, "dur": 1.488,
    "args": {
      "Python parent id": 8166, "Python id": 8167, "finished": true, "Ev Idx": 11067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.469, "dur": 1.130,
    "args": {
      "Python parent id": 8167, "Python id": 8168, "finished": true, "Ev Idx": 11068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.648, "dur": 0.472,
    "args": {
      "Python parent id": 8168, "Python id": 8169, "finished": true, "Ev Idx": 11069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.730, "dur": 0.055,
    "args": {
      "Python parent id": 8169, "Python id": 8170, "finished": true, "Ev Idx": 11070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910385.929, "dur": 0.147,
    "args": {
      "Python parent id": 8169, "Python id": 8171, "finished": true, "Ev Idx": 11071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910387.156, "dur": 0.092,
    "args": {
      "Python parent id": 8162, "Python id": 8172, "finished": true, "Ev Idx": 11072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910387.527, "dur": 0.068,
    "args": {
      "Python parent id": 8160, "Python id": 8173, "finished": true, "Ev Idx": 11073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910387.842, "dur": 1.154,
    "args": {
      "Python parent id": 8160, "Python id": 8174, "finished": true, "Ev Idx": 11074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910388.024, "dur": 0.934,
    "args": {
      "Python parent id": 8174, "Python id": 8175, "finished": true, "Ev Idx": 11075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910388.164, "dur": 0.361,
    "args": {
      "Python parent id": 8175, "Python id": 8176, "finished": true, "Ev Idx": 11076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910388.297, "dur": 0.169,
    "args": {
      "Python parent id": 8176, "Python id": 8177, "finished": true, "Ev Idx": 11077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910388.387, "dur": 0.049,
    "args": {
      "Python parent id": 8177, "Python id": 8178, "finished": true, "Ev Idx": 11078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910388.642, "dur": 0.272,
    "args": {
      "Python parent id": 8175, "Python id": 8179, "finished": true, "Ev Idx": 11079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910389.350, "dur": 0.133,
    "args": {
      "Python parent id": 8144, "Python id": 8180, "finished": true, "Ev Idx": 11080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910389.997, "dur": 0.886,
    "args": {
      "Python parent id": 8144, "Python id": 8181, "finished": true, "Ev Idx": 11081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910390.386, "dur": 0.448,
    "args": {
      "Python parent id": 8181, "Python id": 8182, "finished": true, "Ev Idx": 11082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910391.242, "dur": 3.383,
    "args": {
      "Python parent id": 8143, "Python id": 8183, "finished": true, "Ev Idx": 11083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910395.579, "dur": 90.925,
    "args": {
      "Python parent id": 12, "Python id": 8184, "finished": true, "Ev Idx": 11084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910395.901, "dur": 0.529,
    "args": {
      "Python parent id": 8184, "Python id": 8185, "finished": true, "Ev Idx": 11085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910396.220, "dur": 0.133,
    "args": {
      "Python parent id": 8185, "Python id": 8186, "finished": true, "Ev Idx": 11086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910396.729, "dur": 0.393,
    "args": {
      "Python parent id": 8184, "Python id": 8187, "finished": true, "Ev Idx": 11087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910397.297, "dur": 0.756,
    "args": {
      "Python parent id": 8184, "Python id": 8188, "finished": true, "Ev Idx": 11088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910398.274, "dur": 1.570,
    "args": {
      "Python parent id": 8184, "Python id": 8189, "finished": true, "Ev Idx": 11089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910400.453, "dur": 85.985,
    "args": {
      "Python parent id": 8184, "Python id": 8190, "finished": true, "Ev Idx": 11090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910401.333, "dur": 84.898,
    "args": {
      "Python parent id": 8190, "Python id": 8191, "finished": true, "Ev Idx": 11091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910401.646, "dur": 84.531,
    "args": {
      "Python parent id": 8191, "Python id": 8192, "finished": true, "Ev Idx": 11092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910406.697, "dur": 78.102,
    "args": {
      "Python parent id": 8192, "Python id": 8193, "finished": true, "Ev Idx": 11093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910407.019, "dur": 77.687,
    "args": {
      "Python parent id": 8193, "Python id": 8194, "finished": true, "Ev Idx": 11094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910407.604, "dur": 77.040,
    "args": {
      "Python parent id": 8194, "Python id": 8195, "finished": true, "Ev Idx": 11095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910408.138, "dur": 0.403,
    "args": {
      "Python parent id": 8195, "Python id": 8196, "finished": true, "Ev Idx": 11096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910408.430, "dur": 0.058,
    "args": {
      "Python parent id": 8196, "Python id": 8197, "finished": true, "Ev Idx": 11097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910408.870, "dur": 0.307,
    "args": {
      "Python parent id": 8195, "Python id": 8198, "finished": true, "Ev Idx": 11098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910409.094, "dur": 0.032,
    "args": {
      "Python parent id": 8198, "Python id": 8199, "finished": true, "Ev Idx": 11099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910409.546, "dur": 75.002,
    "args": {
      "Python parent id": 8195, "Python id": 8200, "finished": true, "Ev Idx": 11100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910410.256, "dur": 73.023,
    "args": {
      "Python parent id": 8200, "Python id": 8201, "finished": true, "Ev Idx": 11101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910417.051, "dur": 18.382,
    "args": {
      "Python parent id": 8201, "Python id": 8202, "finished": true, "Ev Idx": 11102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910417.198, "dur": 0.142,
    "args": {
      "Python parent id": 8202, "Python id": 8203, "finished": true, "Ev Idx": 11103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910417.669, "dur": 17.609,
    "args": {
      "Python parent id": 8202, "Python id": 8204, "finished": true, "Ev Idx": 11104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910417.790, "dur": 8.330,
    "args": {
      "Python parent id": 8204, "Python id": 8205, "finished": true, "Ev Idx": 11105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910444.496, "dur": 7.530,
    "args": {
      "Python parent id": 8201, "Python id": 8206, "finished": true, "Ev Idx": 11106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910461.127, "dur": 7.811,
    "args": {
      "Python parent id": 8201, "Python id": 8207, "finished": true, "Ev Idx": 11107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910487.772, "dur": 17.162,
    "args": {
      "Python parent id": 12, "Python id": 8208, "finished": true, "Ev Idx": 11108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910488.144, "dur": 12.753,
    "args": {
      "Python parent id": 8208, "Python id": 8209, "finished": true, "Ev Idx": 11109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910488.332, "dur": 0.800,
    "args": {
      "Python parent id": 8209, "Python id": 8210, "finished": true, "Ev Idx": 11110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910488.747, "dur": 0.320,
    "args": {
      "Python parent id": 8210, "Python id": 8211, "finished": true, "Ev Idx": 11111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910488.927, "dur": 0.109,
    "args": {
      "Python parent id": 8211, "Python id": 8212, "finished": true, "Ev Idx": 11112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910489.492, "dur": 9.197,
    "args": {
      "Python parent id": 8209, "Python id": 8213, "finished": true, "Ev Idx": 11113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910489.608, "dur": 0.068,
    "args": {
      "Python parent id": 8213, "Python id": 8214, "finished": true, "Ev Idx": 11114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910489.802, "dur": 0.056,
    "args": {
      "Python parent id": 8213, "Python id": 8215, "finished": true, "Ev Idx": 11115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910489.987, "dur": 0.054,
    "args": {
      "Python parent id": 8213, "Python id": 8216, "finished": true, "Ev Idx": 11116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910490.476, "dur": 0.052,
    "args": {
      "Python parent id": 8213, "Python id": 8217, "finished": true, "Ev Idx": 11117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910491.081, "dur": 0.061,
    "args": {
      "Python parent id": 8213, "Python id": 8218, "finished": true, "Ev Idx": 11118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910491.492, "dur": 7.123,
    "args": {
      "Python parent id": 8213, "Python id": 8219, "finished": true, "Ev Idx": 11119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910491.603, "dur": 0.057,
    "args": {
      "Python parent id": 8219, "Python id": 8220, "finished": true, "Ev Idx": 11120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910491.808, "dur": 0.059,
    "args": {
      "Python parent id": 8219, "Python id": 8221, "finished": true, "Ev Idx": 11121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910492.040, "dur": 0.079,
    "args": {
      "Python parent id": 8219, "Python id": 8222, "finished": true, "Ev Idx": 11122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910492.480, "dur": 0.046,
    "args": {
      "Python parent id": 8219, "Python id": 8223, "finished": true, "Ev Idx": 11123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910492.868, "dur": 5.688,
    "args": {
      "Python parent id": 8219, "Python id": 8224, "finished": true, "Ev Idx": 11124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910493.127, "dur": 5.373,
    "args": {
      "Python parent id": 8224, "Python id": 8225, "finished": true, "Ev Idx": 11125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910493.255, "dur": 3.508,
    "args": {
      "Python parent id": 8225, "Python id": 8226, "finished": true, "Ev Idx": 11126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910493.410, "dur": 3.278,
    "args": {
      "Python parent id": 8226, "Python id": 8227, "finished": true, "Ev Idx": 11127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910493.566, "dur": 0.285,
    "args": {
      "Python parent id": 8227, "Python id": 8228, "finished": true, "Ev Idx": 11128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910493.716, "dur": 0.057,
    "args": {
      "Python parent id": 8228, "Python id": 8229, "finished": true, "Ev Idx": 11129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.068, "dur": 2.229,
    "args": {
      "Python parent id": 8227, "Python id": 8230, "finished": true, "Ev Idx": 11130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.262, "dur": 1.952,
    "args": {
      "Python parent id": 8230, "Python id": 8231, "finished": true, "Ev Idx": 11131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.486, "dur": 1.665,
    "args": {
      "Python parent id": 8231, "Python id": 8232, "finished": true, "Ev Idx": 11132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.671, "dur": 1.226,
    "args": {
      "Python parent id": 8232, "Python id": 8233, "finished": true, "Ev Idx": 11133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.843, "dur": 0.498,
    "args": {
      "Python parent id": 8233, "Python id": 8234, "finished": true, "Ev Idx": 11134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910494.936, "dur": 0.046,
    "args": {
      "Python parent id": 8234, "Python id": 8235, "finished": true, "Ev Idx": 11135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910495.152, "dur": 0.138,
    "args": {
      "Python parent id": 8234, "Python id": 8236, "finished": true, "Ev Idx": 11136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910496.494, "dur": 0.153,
    "args": {
      "Python parent id": 8227, "Python id": 8237, "finished": true, "Ev Idx": 11137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910496.915, "dur": 0.069,
    "args": {
      "Python parent id": 8225, "Python id": 8238, "finished": true, "Ev Idx": 11138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910497.285, "dur": 1.151,
    "args": {
      "Python parent id": 8225, "Python id": 8239, "finished": true, "Ev Idx": 11139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910497.411, "dur": 0.965,
    "args": {
      "Python parent id": 8239, "Python id": 8240, "finished": true, "Ev Idx": 11140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910497.564, "dur": 0.370,
    "args": {
      "Python parent id": 8240, "Python id": 8241, "finished": true, "Ev Idx": 11141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910497.689, "dur": 0.185,
    "args": {
      "Python parent id": 8241, "Python id": 8242, "finished": true, "Ev Idx": 11142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910497.789, "dur": 0.052,
    "args": {
      "Python parent id": 8242, "Python id": 8243, "finished": true, "Ev Idx": 11143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910498.059, "dur": 0.289,
    "args": {
      "Python parent id": 8240, "Python id": 8244, "finished": true, "Ev Idx": 11144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910498.811, "dur": 0.221,
    "args": {
      "Python parent id": 8209, "Python id": 8245, "finished": true, "Ev Idx": 11145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910499.609, "dur": 1.131,
    "args": {
      "Python parent id": 8209, "Python id": 8246, "finished": true, "Ev Idx": 11146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910500.107, "dur": 0.563,
    "args": {
      "Python parent id": 8246, "Python id": 8247, "finished": true, "Ev Idx": 11147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910501.186, "dur": 3.642,
    "args": {
      "Python parent id": 8208, "Python id": 8248, "finished": true, "Ev Idx": 11148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910505.542, "dur": 6.856,
    "args": {
      "Python parent id": 12, "Python id": 8249, "finished": true, "Ev Idx": 11149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910512.842, "dur": 0.090,
    "args": {
      "Python parent id": 12, "Python id": 8250, "finished": true, "Ev Idx": 11150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910513.545, "dur": 0.180,
    "args": {
      "Python parent id": 12, "Python id": 8251, "finished": true, "Ev Idx": 11151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910514.197, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 8252, "finished": true, "Ev Idx": 11152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910514.670, "dur": 9.909,
    "args": {
      "Python parent id": 12, "Python id": 8253, "finished": true, "Ev Idx": 11153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910525.968, "dur": 14.851,
    "args": {
      "Python parent id": 12, "Python id": 8254, "finished": true, "Ev Idx": 11154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910526.242, "dur": 10.429,
    "args": {
      "Python parent id": 8254, "Python id": 8255, "finished": true, "Ev Idx": 11155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910526.387, "dur": 0.457,
    "args": {
      "Python parent id": 8255, "Python id": 8256, "finished": true, "Ev Idx": 11156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910526.550, "dur": 0.229,
    "args": {
      "Python parent id": 8256, "Python id": 8257, "finished": true, "Ev Idx": 11157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910526.634, "dur": 0.113,
    "args": {
      "Python parent id": 8257, "Python id": 8258, "finished": true, "Ev Idx": 11158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910527.082, "dur": 7.827,
    "args": {
      "Python parent id": 8255, "Python id": 8259, "finished": true, "Ev Idx": 11159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910527.161, "dur": 0.064,
    "args": {
      "Python parent id": 8259, "Python id": 8260, "finished": true, "Ev Idx": 11160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910527.328, "dur": 0.057,
    "args": {
      "Python parent id": 8259, "Python id": 8261, "finished": true, "Ev Idx": 11161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910527.535, "dur": 0.059,
    "args": {
      "Python parent id": 8259, "Python id": 8262, "finished": true, "Ev Idx": 11162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910527.934, "dur": 0.042,
    "args": {
      "Python parent id": 8259, "Python id": 8263, "finished": true, "Ev Idx": 11163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910528.239, "dur": 0.071,
    "args": {
      "Python parent id": 8259, "Python id": 8264, "finished": true, "Ev Idx": 11164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910528.626, "dur": 6.228,
    "args": {
      "Python parent id": 8259, "Python id": 8265, "finished": true, "Ev Idx": 11165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910528.714, "dur": 0.059,
    "args": {
      "Python parent id": 8265, "Python id": 8266, "finished": true, "Ev Idx": 11166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910528.950, "dur": 0.041,
    "args": {
      "Python parent id": 8265, "Python id": 8267, "finished": true, "Ev Idx": 11167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910529.169, "dur": 0.063,
    "args": {
      "Python parent id": 8265, "Python id": 8268, "finished": true, "Ev Idx": 11168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910529.564, "dur": 0.031,
    "args": {
      "Python parent id": 8265, "Python id": 8269, "finished": true, "Ev Idx": 11169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910529.896, "dur": 4.871,
    "args": {
      "Python parent id": 8265, "Python id": 8270, "finished": true, "Ev Idx": 11170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.082, "dur": 4.631,
    "args": {
      "Python parent id": 8270, "Python id": 8271, "finished": true, "Ev Idx": 11171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.199, "dur": 2.871,
    "args": {
      "Python parent id": 8271, "Python id": 8272, "finished": true, "Ev Idx": 11172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.369, "dur": 2.658,
    "args": {
      "Python parent id": 8272, "Python id": 8273, "finished": true, "Ev Idx": 11173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.499, "dur": 0.275,
    "args": {
      "Python parent id": 8273, "Python id": 8274, "finished": true, "Ev Idx": 11174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.641, "dur": 0.057,
    "args": {
      "Python parent id": 8274, "Python id": 8275, "finished": true, "Ev Idx": 11175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910530.957, "dur": 1.747,
    "args": {
      "Python parent id": 8273, "Python id": 8276, "finished": true, "Ev Idx": 11176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.140, "dur": 1.507,
    "args": {
      "Python parent id": 8276, "Python id": 8277, "finished": true, "Ev Idx": 11177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.315, "dur": 1.285,
    "args": {
      "Python parent id": 8277, "Python id": 8278, "finished": true, "Ev Idx": 11178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.427, "dur": 0.993,
    "args": {
      "Python parent id": 8278, "Python id": 8279, "finished": true, "Ev Idx": 11179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.567, "dur": 0.453,
    "args": {
      "Python parent id": 8279, "Python id": 8280, "finished": true, "Ev Idx": 11180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.634, "dur": 0.045,
    "args": {
      "Python parent id": 8280, "Python id": 8281, "finished": true, "Ev Idx": 11181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910531.820, "dur": 0.150,
    "args": {
      "Python parent id": 8280, "Python id": 8282, "finished": true, "Ev Idx": 11182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910532.884, "dur": 0.094,
    "args": {
      "Python parent id": 8273, "Python id": 8283, "finished": true, "Ev Idx": 11183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910533.232, "dur": 0.089,
    "args": {
      "Python parent id": 8271, "Python id": 8284, "finished": true, "Ev Idx": 11184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910533.550, "dur": 1.100,
    "args": {
      "Python parent id": 8271, "Python id": 8285, "finished": true, "Ev Idx": 11185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910533.676, "dur": 0.932,
    "args": {
      "Python parent id": 8285, "Python id": 8286, "finished": true, "Ev Idx": 11186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910533.810, "dur": 0.377,
    "args": {
      "Python parent id": 8286, "Python id": 8287, "finished": true, "Ev Idx": 11187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910533.936, "dur": 0.192,
    "args": {
      "Python parent id": 8287, "Python id": 8288, "finished": true, "Ev Idx": 11188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910534.038, "dur": 0.057,
    "args": {
      "Python parent id": 8288, "Python id": 8289, "finished": true, "Ev Idx": 11189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910534.313, "dur": 0.264,
    "args": {
      "Python parent id": 8286, "Python id": 8290, "finished": true, "Ev Idx": 11190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910535.022, "dur": 0.175,
    "args": {
      "Python parent id": 8255, "Python id": 8291, "finished": true, "Ev Idx": 11191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910535.739, "dur": 0.813,
    "args": {
      "Python parent id": 8255, "Python id": 8292, "finished": true, "Ev Idx": 11192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910536.077, "dur": 0.425,
    "args": {
      "Python parent id": 8292, "Python id": 8293, "finished": true, "Ev Idx": 11193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910536.937, "dur": 3.832,
    "args": {
      "Python parent id": 8254, "Python id": 8294, "finished": true, "Ev Idx": 11194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910541.655, "dur": 111.790,
    "args": {
      "Python parent id": 12, "Python id": 8295, "finished": true, "Ev Idx": 11195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910542.051, "dur": 0.445,
    "args": {
      "Python parent id": 8295, "Python id": 8296, "finished": true, "Ev Idx": 11196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910542.313, "dur": 0.119,
    "args": {
      "Python parent id": 8296, "Python id": 8297, "finished": true, "Ev Idx": 11197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910542.870, "dur": 0.349,
    "args": {
      "Python parent id": 8295, "Python id": 8298, "finished": true, "Ev Idx": 11198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910543.430, "dur": 0.272,
    "args": {
      "Python parent id": 8295, "Python id": 8299, "finished": true, "Ev Idx": 11199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910543.907, "dur": 1.721,
    "args": {
      "Python parent id": 8295, "Python id": 8300, "finished": true, "Ev Idx": 11200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910546.269, "dur": 107.105,
    "args": {
      "Python parent id": 8295, "Python id": 8301, "finished": true, "Ev Idx": 11201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910547.116, "dur": 106.062,
    "args": {
      "Python parent id": 8301, "Python id": 8302, "finished": true, "Ev Idx": 11202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910547.445, "dur": 105.630,
    "args": {
      "Python parent id": 8302, "Python id": 8303, "finished": true, "Ev Idx": 11203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910553.013, "dur": 98.536,
    "args": {
      "Python parent id": 8303, "Python id": 8304, "finished": true, "Ev Idx": 11204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910553.395, "dur": 98.065,
    "args": {
      "Python parent id": 8304, "Python id": 8305, "finished": true, "Ev Idx": 11205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910553.980, "dur": 97.302,
    "args": {
      "Python parent id": 8305, "Python id": 8306, "finished": true, "Ev Idx": 11206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910554.587, "dur": 0.412,
    "args": {
      "Python parent id": 8306, "Python id": 8307, "finished": true, "Ev Idx": 11207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910554.877, "dur": 0.045,
    "args": {
      "Python parent id": 8307, "Python id": 8308, "finished": true, "Ev Idx": 11208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910555.313, "dur": 0.344,
    "args": {
      "Python parent id": 8306, "Python id": 8309, "finished": true, "Ev Idx": 11209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910555.565, "dur": 0.038,
    "args": {
      "Python parent id": 8309, "Python id": 8310, "finished": true, "Ev Idx": 11210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910555.980, "dur": 95.209,
    "args": {
      "Python parent id": 8306, "Python id": 8311, "finished": true, "Ev Idx": 11211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910556.729, "dur": 93.193,
    "args": {
      "Python parent id": 8311, "Python id": 8312, "finished": true, "Ev Idx": 11212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910572.628, "dur": 19.452,
    "args": {
      "Python parent id": 8312, "Python id": 8313, "finished": true, "Ev Idx": 11213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910572.818, "dur": 0.126,
    "args": {
      "Python parent id": 8313, "Python id": 8314, "finished": true, "Ev Idx": 11214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910573.295, "dur": 18.652,
    "args": {
      "Python parent id": 8313, "Python id": 8315, "finished": true, "Ev Idx": 11215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910573.398, "dur": 8.334,
    "args": {
      "Python parent id": 8315, "Python id": 8316, "finished": true, "Ev Idx": 11216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910599.584, "dur": 11.864,
    "args": {
      "Python parent id": 8312, "Python id": 8317, "finished": true, "Ev Idx": 11217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910624.662, "dur": 9.292,
    "args": {
      "Python parent id": 8312, "Python id": 8318, "finished": true, "Ev Idx": 11218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910654.819, "dur": 18.579,
    "args": {
      "Python parent id": 12, "Python id": 8319, "finished": true, "Ev Idx": 11219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910655.209, "dur": 13.812,
    "args": {
      "Python parent id": 8319, "Python id": 8320, "finished": true, "Ev Idx": 11220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910655.472, "dur": 0.495,
    "args": {
      "Python parent id": 8320, "Python id": 8321, "finished": true, "Ev Idx": 11221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910655.614, "dur": 0.282,
    "args": {
      "Python parent id": 8321, "Python id": 8322, "finished": true, "Ev Idx": 11222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910655.783, "dur": 0.076,
    "args": {
      "Python parent id": 8322, "Python id": 8323, "finished": true, "Ev Idx": 11223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910656.366, "dur": 10.135,
    "args": {
      "Python parent id": 8320, "Python id": 8324, "finished": true, "Ev Idx": 11224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910656.524, "dur": 0.116,
    "args": {
      "Python parent id": 8324, "Python id": 8325, "finished": true, "Ev Idx": 11225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910656.763, "dur": 0.068,
    "args": {
      "Python parent id": 8324, "Python id": 8326, "finished": true, "Ev Idx": 11226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910657.010, "dur": 0.062,
    "args": {
      "Python parent id": 8324, "Python id": 8327, "finished": true, "Ev Idx": 11227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910657.486, "dur": 0.039,
    "args": {
      "Python parent id": 8324, "Python id": 8328, "finished": true, "Ev Idx": 11228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910657.817, "dur": 0.106,
    "args": {
      "Python parent id": 8324, "Python id": 8329, "finished": true, "Ev Idx": 11229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910658.361, "dur": 8.078,
    "args": {
      "Python parent id": 8324, "Python id": 8330, "finished": true, "Ev Idx": 11230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910658.470, "dur": 0.066,
    "args": {
      "Python parent id": 8330, "Python id": 8331, "finished": true, "Ev Idx": 11231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910658.716, "dur": 0.065,
    "args": {
      "Python parent id": 8330, "Python id": 8332, "finished": true, "Ev Idx": 11232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910659.033, "dur": 0.069,
    "args": {
      "Python parent id": 8330, "Python id": 8333, "finished": true, "Ev Idx": 11233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910659.508, "dur": 0.070,
    "args": {
      "Python parent id": 8330, "Python id": 8334, "finished": true, "Ev Idx": 11234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910659.921, "dur": 6.407,
    "args": {
      "Python parent id": 8330, "Python id": 8335, "finished": true, "Ev Idx": 11235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910660.219, "dur": 6.043,
    "args": {
      "Python parent id": 8335, "Python id": 8336, "finished": true, "Ev Idx": 11236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910660.384, "dur": 3.872,
    "args": {
      "Python parent id": 8336, "Python id": 8337, "finished": true, "Ev Idx": 11237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910660.579, "dur": 3.579,
    "args": {
      "Python parent id": 8337, "Python id": 8338, "finished": true, "Ev Idx": 11238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910660.703, "dur": 0.286,
    "args": {
      "Python parent id": 8338, "Python id": 8339, "finished": true, "Ev Idx": 11239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910660.865, "dur": 0.064,
    "args": {
      "Python parent id": 8339, "Python id": 8340, "finished": true, "Ev Idx": 11240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910661.153, "dur": 2.613,
    "args": {
      "Python parent id": 8338, "Python id": 8341, "finished": true, "Ev Idx": 11241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910661.416, "dur": 2.232,
    "args": {
      "Python parent id": 8341, "Python id": 8342, "finished": true, "Ev Idx": 11242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910661.637, "dur": 1.942,
    "args": {
      "Python parent id": 8342, "Python id": 8343, "finished": true, "Ev Idx": 11243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910661.810, "dur": 1.463,
    "args": {
      "Python parent id": 8343, "Python id": 8344, "finished": true, "Ev Idx": 11244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910662.048, "dur": 0.638,
    "args": {
      "Python parent id": 8344, "Python id": 8345, "finished": true, "Ev Idx": 11245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910662.187, "dur": 0.066,
    "args": {
      "Python parent id": 8345, "Python id": 8346, "finished": true, "Ev Idx": 11246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910662.463, "dur": 0.166,
    "args": {
      "Python parent id": 8345, "Python id": 8347, "finished": true, "Ev Idx": 11247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910663.969, "dur": 0.149,
    "args": {
      "Python parent id": 8338, "Python id": 8348, "finished": true, "Ev Idx": 11248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910664.483, "dur": 0.107,
    "args": {
      "Python parent id": 8336, "Python id": 8349, "finished": true, "Ev Idx": 11249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910664.887, "dur": 1.304,
    "args": {
      "Python parent id": 8336, "Python id": 8350, "finished": true, "Ev Idx": 11250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910665.043, "dur": 1.110,
    "args": {
      "Python parent id": 8350, "Python id": 8351, "finished": true, "Ev Idx": 11251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910665.227, "dur": 0.398,
    "args": {
      "Python parent id": 8351, "Python id": 8352, "finished": true, "Ev Idx": 11252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910665.385, "dur": 0.175,
    "args": {
      "Python parent id": 8352, "Python id": 8353, "finished": true, "Ev Idx": 11253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910665.469, "dur": 0.055,
    "args": {
      "Python parent id": 8353, "Python id": 8354, "finished": true, "Ev Idx": 11254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910665.763, "dur": 0.335,
    "args": {
      "Python parent id": 8351, "Python id": 8355, "finished": true, "Ev Idx": 11255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910666.631, "dur": 0.263,
    "args": {
      "Python parent id": 8320, "Python id": 8356, "finished": true, "Ev Idx": 11256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910667.637, "dur": 1.234,
    "args": {
      "Python parent id": 8320, "Python id": 8357, "finished": true, "Ev Idx": 11257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910668.237, "dur": 0.541,
    "args": {
      "Python parent id": 8357, "Python id": 8358, "finished": true, "Ev Idx": 11258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910669.350, "dur": 3.959,
    "args": {
      "Python parent id": 8319, "Python id": 8359, "finished": true, "Ev Idx": 11259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910674.022, "dur": 8.042,
    "args": {
      "Python parent id": 12, "Python id": 8360, "finished": true, "Ev Idx": 11260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910682.663, "dur": 0.114,
    "args": {
      "Python parent id": 12, "Python id": 8361, "finished": true, "Ev Idx": 11261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910683.430, "dur": 0.237,
    "args": {
      "Python parent id": 12, "Python id": 8362, "finished": true, "Ev Idx": 11262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910684.165, "dur": 0.078,
    "args": {
      "Python parent id": 12, "Python id": 8363, "finished": true, "Ev Idx": 11263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910684.650, "dur": 10.664,
    "args": {
      "Python parent id": 12, "Python id": 8364, "finished": true, "Ev Idx": 11264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910696.788, "dur": 16.741,
    "args": {
      "Python parent id": 12, "Python id": 8365, "finished": true, "Ev Idx": 11265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910697.151, "dur": 12.025,
    "args": {
      "Python parent id": 8365, "Python id": 8366, "finished": true, "Ev Idx": 11266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910697.305, "dur": 0.506,
    "args": {
      "Python parent id": 8366, "Python id": 8367, "finished": true, "Ev Idx": 11267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910697.503, "dur": 0.237,
    "args": {
      "Python parent id": 8367, "Python id": 8368, "finished": true, "Ev Idx": 11268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910697.599, "dur": 0.106,
    "args": {
      "Python parent id": 8368, "Python id": 8369, "finished": true, "Ev Idx": 11269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910698.061, "dur": 9.173,
    "args": {
      "Python parent id": 8366, "Python id": 8370, "finished": true, "Ev Idx": 11270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910698.153, "dur": 0.066,
    "args": {
      "Python parent id": 8370, "Python id": 8371, "finished": true, "Ev Idx": 11271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910698.340, "dur": 0.057,
    "args": {
      "Python parent id": 8370, "Python id": 8372, "finished": true, "Ev Idx": 11272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910698.551, "dur": 0.059,
    "args": {
      "Python parent id": 8370, "Python id": 8373, "finished": true, "Ev Idx": 11273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910699.055, "dur": 0.042,
    "args": {
      "Python parent id": 8370, "Python id": 8374, "finished": true, "Ev Idx": 11274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910699.452, "dur": 0.087,
    "args": {
      "Python parent id": 8370, "Python id": 8375, "finished": true, "Ev Idx": 11275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910699.965, "dur": 7.191,
    "args": {
      "Python parent id": 8370, "Python id": 8376, "finished": true, "Ev Idx": 11276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910700.060, "dur": 0.062,
    "args": {
      "Python parent id": 8376, "Python id": 8377, "finished": true, "Ev Idx": 11277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910700.286, "dur": 0.058,
    "args": {
      "Python parent id": 8376, "Python id": 8378, "finished": true, "Ev Idx": 11278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910700.569, "dur": 0.072,
    "args": {
      "Python parent id": 8376, "Python id": 8379, "finished": true, "Ev Idx": 11279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910700.981, "dur": 0.072,
    "args": {
      "Python parent id": 8376, "Python id": 8380, "finished": true, "Ev Idx": 11280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910701.386, "dur": 5.642,
    "args": {
      "Python parent id": 8376, "Python id": 8381, "finished": true, "Ev Idx": 11281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910701.674, "dur": 5.295,
    "args": {
      "Python parent id": 8381, "Python id": 8382, "finished": true, "Ev Idx": 11282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910701.801, "dur": 3.316,
    "args": {
      "Python parent id": 8382, "Python id": 8383, "finished": true, "Ev Idx": 11283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910702.018, "dur": 3.024,
    "args": {
      "Python parent id": 8383, "Python id": 8384, "finished": true, "Ev Idx": 11284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910702.160, "dur": 0.280,
    "args": {
      "Python parent id": 8384, "Python id": 8385, "finished": true, "Ev Idx": 11285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910702.313, "dur": 0.073,
    "args": {
      "Python parent id": 8385, "Python id": 8386, "finished": true, "Ev Idx": 11286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910702.611, "dur": 2.100,
    "args": {
      "Python parent id": 8384, "Python id": 8387, "finished": true, "Ev Idx": 11287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910702.819, "dur": 1.818,
    "args": {
      "Python parent id": 8387, "Python id": 8388, "finished": true, "Ev Idx": 11288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910703.002, "dur": 1.584,
    "args": {
      "Python parent id": 8388, "Python id": 8389, "finished": true, "Ev Idx": 11289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910703.160, "dur": 1.194,
    "args": {
      "Python parent id": 8389, "Python id": 8390, "finished": true, "Ev Idx": 11290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910703.363, "dur": 0.550,
    "args": {
      "Python parent id": 8390, "Python id": 8391, "finished": true, "Ev Idx": 11291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910703.490, "dur": 0.041,
    "args": {
      "Python parent id": 8391, "Python id": 8392, "finished": true, "Ev Idx": 11292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910703.730, "dur": 0.126,
    "args": {
      "Python parent id": 8391, "Python id": 8393, "finished": true, "Ev Idx": 11293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910704.887, "dur": 0.117,
    "args": {
      "Python parent id": 8384, "Python id": 8394, "finished": true, "Ev Idx": 11294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910705.326, "dur": 0.099,
    "args": {
      "Python parent id": 8382, "Python id": 8395, "finished": true, "Ev Idx": 11295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910705.702, "dur": 1.222,
    "args": {
      "Python parent id": 8382, "Python id": 8396, "finished": true, "Ev Idx": 11296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910705.851, "dur": 1.029,
    "args": {
      "Python parent id": 8396, "Python id": 8397, "finished": true, "Ev Idx": 11297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910705.998, "dur": 0.429,
    "args": {
      "Python parent id": 8397, "Python id": 8398, "finished": true, "Ev Idx": 11298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910706.193, "dur": 0.173,
    "args": {
      "Python parent id": 8398, "Python id": 8399, "finished": true, "Ev Idx": 11299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910706.267, "dur": 0.063,
    "args": {
      "Python parent id": 8399, "Python id": 8400, "finished": true, "Ev Idx": 11300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910706.562, "dur": 0.291,
    "args": {
      "Python parent id": 8397, "Python id": 8401, "finished": true, "Ev Idx": 11301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910707.377, "dur": 0.170,
    "args": {
      "Python parent id": 8366, "Python id": 8402, "finished": true, "Ev Idx": 11302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910708.131, "dur": 0.918,
    "args": {
      "Python parent id": 8366, "Python id": 8403, "finished": true, "Ev Idx": 11303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910708.503, "dur": 0.493,
    "args": {
      "Python parent id": 8403, "Python id": 8404, "finished": true, "Ev Idx": 11304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910709.461, "dur": 4.008,
    "args": {
      "Python parent id": 8365, "Python id": 8405, "finished": true, "Ev Idx": 11305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910714.434, "dur": 96.281,
    "args": {
      "Python parent id": 12, "Python id": 8406, "finished": true, "Ev Idx": 11306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910714.815, "dur": 0.455,
    "args": {
      "Python parent id": 8406, "Python id": 8407, "finished": true, "Ev Idx": 11307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910715.090, "dur": 0.116,
    "args": {
      "Python parent id": 8407, "Python id": 8408, "finished": true, "Ev Idx": 11308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910715.645, "dur": 0.367,
    "args": {
      "Python parent id": 8406, "Python id": 8409, "finished": true, "Ev Idx": 11309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910716.245, "dur": 0.244,
    "args": {
      "Python parent id": 8406, "Python id": 8410, "finished": true, "Ev Idx": 11310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910716.709, "dur": 1.703,
    "args": {
      "Python parent id": 8406, "Python id": 8411, "finished": true, "Ev Idx": 11311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910719.048, "dur": 91.596,
    "args": {
      "Python parent id": 8406, "Python id": 8412, "finished": true, "Ev Idx": 11312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910719.912, "dur": 90.511,
    "args": {
      "Python parent id": 8412, "Python id": 8413, "finished": true, "Ev Idx": 11313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910720.231, "dur": 90.114,
    "args": {
      "Python parent id": 8413, "Python id": 8414, "finished": true, "Ev Idx": 11314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910725.642, "dur": 83.319,
    "args": {
      "Python parent id": 8414, "Python id": 8415, "finished": true, "Ev Idx": 11315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910725.986, "dur": 82.801,
    "args": {
      "Python parent id": 8415, "Python id": 8416, "finished": true, "Ev Idx": 11316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910726.594, "dur": 82.126,
    "args": {
      "Python parent id": 8416, "Python id": 8417, "finished": true, "Ev Idx": 11317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910727.130, "dur": 0.446,
    "args": {
      "Python parent id": 8417, "Python id": 8418, "finished": true, "Ev Idx": 11318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910727.436, "dur": 0.090,
    "args": {
      "Python parent id": 8418, "Python id": 8419, "finished": true, "Ev Idx": 11319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910727.902, "dur": 0.387,
    "args": {
      "Python parent id": 8417, "Python id": 8420, "finished": true, "Ev Idx": 11320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910728.188, "dur": 0.040,
    "args": {
      "Python parent id": 8420, "Python id": 8421, "finished": true, "Ev Idx": 11321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910728.608, "dur": 79.990,
    "args": {
      "Python parent id": 8417, "Python id": 8422, "finished": true, "Ev Idx": 11322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910729.287, "dur": 78.009,
    "args": {
      "Python parent id": 8422, "Python id": 8423, "finished": true, "Ev Idx": 11323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910737.517, "dur": 18.906,
    "args": {
      "Python parent id": 8423, "Python id": 8424, "finished": true, "Ev Idx": 11324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910737.677, "dur": 0.132,
    "args": {
      "Python parent id": 8424, "Python id": 8425, "finished": true, "Ev Idx": 11325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910738.173, "dur": 18.100,
    "args": {
      "Python parent id": 8424, "Python id": 8426, "finished": true, "Ev Idx": 11326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910738.299, "dur": 8.521,
    "args": {
      "Python parent id": 8426, "Python id": 8427, "finished": true, "Ev Idx": 11327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910765.866, "dur": 7.258,
    "args": {
      "Python parent id": 8423, "Python id": 8428, "finished": true, "Ev Idx": 11328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910782.515, "dur": 8.478,
    "args": {
      "Python parent id": 8423, "Python id": 8429, "finished": true, "Ev Idx": 11329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910811.954, "dur": 17.990,
    "args": {
      "Python parent id": 12, "Python id": 8430, "finished": true, "Ev Idx": 11330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910812.300, "dur": 13.776,
    "args": {
      "Python parent id": 8430, "Python id": 8431, "finished": true, "Ev Idx": 11331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910812.511, "dur": 0.566,
    "args": {
      "Python parent id": 8431, "Python id": 8432, "finished": true, "Ev Idx": 11332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910812.733, "dur": 0.278,
    "args": {
      "Python parent id": 8432, "Python id": 8433, "finished": true, "Ev Idx": 11333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910812.872, "dur": 0.104,
    "args": {
      "Python parent id": 8433, "Python id": 8434, "finished": true, "Ev Idx": 11334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910813.429, "dur": 10.140,
    "args": {
      "Python parent id": 8431, "Python id": 8435, "finished": true, "Ev Idx": 11335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910813.538, "dur": 0.114,
    "args": {
      "Python parent id": 8435, "Python id": 8436, "finished": true, "Ev Idx": 11336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910813.800, "dur": 0.062,
    "args": {
      "Python parent id": 8435, "Python id": 8437, "finished": true, "Ev Idx": 11337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910814.067, "dur": 0.063,
    "args": {
      "Python parent id": 8435, "Python id": 8438, "finished": true, "Ev Idx": 11338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910814.642, "dur": 0.047,
    "args": {
      "Python parent id": 8435, "Python id": 8439, "finished": true, "Ev Idx": 11339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910815.024, "dur": 0.093,
    "args": {
      "Python parent id": 8435, "Python id": 8440, "finished": true, "Ev Idx": 11340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910815.516, "dur": 7.999,
    "args": {
      "Python parent id": 8435, "Python id": 8441, "finished": true, "Ev Idx": 11341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910815.646, "dur": 0.062,
    "args": {
      "Python parent id": 8441, "Python id": 8442, "finished": true, "Ev Idx": 11342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910815.930, "dur": 0.059,
    "args": {
      "Python parent id": 8441, "Python id": 8443, "finished": true, "Ev Idx": 11343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910816.164, "dur": 0.068,
    "args": {
      "Python parent id": 8441, "Python id": 8444, "finished": true, "Ev Idx": 11344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910816.656, "dur": 0.045,
    "args": {
      "Python parent id": 8441, "Python id": 8445, "finished": true, "Ev Idx": 11345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910817.039, "dur": 6.358,
    "args": {
      "Python parent id": 8441, "Python id": 8446, "finished": true, "Ev Idx": 11346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910817.341, "dur": 5.966,
    "args": {
      "Python parent id": 8446, "Python id": 8447, "finished": true, "Ev Idx": 11347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910817.511, "dur": 3.783,
    "args": {
      "Python parent id": 8447, "Python id": 8448, "finished": true, "Ev Idx": 11348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910817.784, "dur": 3.431,
    "args": {
      "Python parent id": 8448, "Python id": 8449, "finished": true, "Ev Idx": 11349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910817.948, "dur": 0.307,
    "args": {
      "Python parent id": 8449, "Python id": 8450, "finished": true, "Ev Idx": 11350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910818.092, "dur": 0.067,
    "args": {
      "Python parent id": 8450, "Python id": 8451, "finished": true, "Ev Idx": 11351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910818.468, "dur": 2.317,
    "args": {
      "Python parent id": 8449, "Python id": 8452, "finished": true, "Ev Idx": 11352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910818.707, "dur": 1.975,
    "args": {
      "Python parent id": 8452, "Python id": 8453, "finished": true, "Ev Idx": 11353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910818.899, "dur": 1.698,
    "args": {
      "Python parent id": 8453, "Python id": 8454, "finished": true, "Ev Idx": 11354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910819.071, "dur": 1.258,
    "args": {
      "Python parent id": 8454, "Python id": 8455, "finished": true, "Ev Idx": 11355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910819.296, "dur": 0.493,
    "args": {
      "Python parent id": 8455, "Python id": 8456, "finished": true, "Ev Idx": 11356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910819.375, "dur": 0.050,
    "args": {
      "Python parent id": 8456, "Python id": 8457, "finished": true, "Ev Idx": 11357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910819.557, "dur": 0.187,
    "args": {
      "Python parent id": 8456, "Python id": 8458, "finished": true, "Ev Idx": 11358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910820.969, "dur": 0.182,
    "args": {
      "Python parent id": 8449, "Python id": 8459, "finished": true, "Ev Idx": 11359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910821.516, "dur": 0.107,
    "args": {
      "Python parent id": 8447, "Python id": 8460, "finished": true, "Ev Idx": 11360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910821.975, "dur": 1.264,
    "args": {
      "Python parent id": 8447, "Python id": 8461, "finished": true, "Ev Idx": 11361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910822.152, "dur": 1.043,
    "args": {
      "Python parent id": 8461, "Python id": 8462, "finished": true, "Ev Idx": 11362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910822.295, "dur": 0.427,
    "args": {
      "Python parent id": 8462, "Python id": 8463, "finished": true, "Ev Idx": 11363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910822.483, "dur": 0.170,
    "args": {
      "Python parent id": 8463, "Python id": 8464, "finished": true, "Ev Idx": 11364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910822.566, "dur": 0.057,
    "args": {
      "Python parent id": 8464, "Python id": 8465, "finished": true, "Ev Idx": 11365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910822.855, "dur": 0.306,
    "args": {
      "Python parent id": 8462, "Python id": 8466, "finished": true, "Ev Idx": 11366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910823.723, "dur": 0.286,
    "args": {
      "Python parent id": 8431, "Python id": 8467, "finished": true, "Ev Idx": 11367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910824.693, "dur": 1.200,
    "args": {
      "Python parent id": 8431, "Python id": 8468, "finished": true, "Ev Idx": 11368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910825.221, "dur": 0.619,
    "args": {
      "Python parent id": 8468, "Python id": 8469, "finished": true, "Ev Idx": 11369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910826.416, "dur": 3.456,
    "args": {
      "Python parent id": 8430, "Python id": 8470, "finished": true, "Ev Idx": 11370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910830.565, "dur": 8.819,
    "args": {
      "Python parent id": 12, "Python id": 8471, "finished": true, "Ev Idx": 11371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910839.868, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 8472, "finished": true, "Ev Idx": 11372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910840.603, "dur": 0.181,
    "args": {
      "Python parent id": 12, "Python id": 8473, "finished": true, "Ev Idx": 11373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910841.260, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 8474, "finished": true, "Ev Idx": 11374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910841.746, "dur": 11.033,
    "args": {
      "Python parent id": 12, "Python id": 8475, "finished": true, "Ev Idx": 11375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910854.336, "dur": 16.915,
    "args": {
      "Python parent id": 12, "Python id": 8476, "finished": true, "Ev Idx": 11376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910854.666, "dur": 12.083,
    "args": {
      "Python parent id": 8476, "Python id": 8477, "finished": true, "Ev Idx": 11377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910854.862, "dur": 0.535,
    "args": {
      "Python parent id": 8477, "Python id": 8478, "finished": true, "Ev Idx": 11378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910855.053, "dur": 0.246,
    "args": {
      "Python parent id": 8478, "Python id": 8479, "finished": true, "Ev Idx": 11379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910855.161, "dur": 0.103,
    "args": {
      "Python parent id": 8479, "Python id": 8480, "finished": true, "Ev Idx": 11380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910855.755, "dur": 9.041,
    "args": {
      "Python parent id": 8477, "Python id": 8481, "finished": true, "Ev Idx": 11381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910855.850, "dur": 0.078,
    "args": {
      "Python parent id": 8481, "Python id": 8482, "finished": true, "Ev Idx": 11382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910856.055, "dur": 0.082,
    "args": {
      "Python parent id": 8481, "Python id": 8483, "finished": true, "Ev Idx": 11383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910856.286, "dur": 0.064,
    "args": {
      "Python parent id": 8481, "Python id": 8484, "finished": true, "Ev Idx": 11384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910856.766, "dur": 0.051,
    "args": {
      "Python parent id": 8481, "Python id": 8485, "finished": true, "Ev Idx": 11385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910857.166, "dur": 0.075,
    "args": {
      "Python parent id": 8481, "Python id": 8486, "finished": true, "Ev Idx": 11386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910857.560, "dur": 7.180,
    "args": {
      "Python parent id": 8481, "Python id": 8487, "finished": true, "Ev Idx": 11387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910857.658, "dur": 0.062,
    "args": {
      "Python parent id": 8487, "Python id": 8488, "finished": true, "Ev Idx": 11388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910857.911, "dur": 0.057,
    "args": {
      "Python parent id": 8487, "Python id": 8489, "finished": true, "Ev Idx": 11389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910858.174, "dur": 0.074,
    "args": {
      "Python parent id": 8487, "Python id": 8490, "finished": true, "Ev Idx": 11390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910858.605, "dur": 0.074,
    "args": {
      "Python parent id": 8487, "Python id": 8491, "finished": true, "Ev Idx": 11391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.020, "dur": 5.609,
    "args": {
      "Python parent id": 8487, "Python id": 8492, "finished": true, "Ev Idx": 11392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.283, "dur": 5.243,
    "args": {
      "Python parent id": 8492, "Python id": 8493, "finished": true, "Ev Idx": 11393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.439, "dur": 3.280,
    "args": {
      "Python parent id": 8493, "Python id": 8494, "finished": true, "Ev Idx": 11394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.656, "dur": 3.020,
    "args": {
      "Python parent id": 8494, "Python id": 8495, "finished": true, "Ev Idx": 11395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.844, "dur": 0.262,
    "args": {
      "Python parent id": 8495, "Python id": 8496, "finished": true, "Ev Idx": 11396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910859.971, "dur": 0.066,
    "args": {
      "Python parent id": 8496, "Python id": 8497, "finished": true, "Ev Idx": 11397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910860.284, "dur": 2.064,
    "args": {
      "Python parent id": 8495, "Python id": 8498, "finished": true, "Ev Idx": 11398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910860.469, "dur": 1.791,
    "args": {
      "Python parent id": 8498, "Python id": 8499, "finished": true, "Ev Idx": 11399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910860.641, "dur": 1.570,
    "args": {
      "Python parent id": 8499, "Python id": 8500, "finished": true, "Ev Idx": 11400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910860.785, "dur": 1.187,
    "args": {
      "Python parent id": 8500, "Python id": 8501, "finished": true, "Ev Idx": 11401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910861.004, "dur": 0.516,
    "args": {
      "Python parent id": 8501, "Python id": 8502, "finished": true, "Ev Idx": 11402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910861.105, "dur": 0.040,
    "args": {
      "Python parent id": 8502, "Python id": 8503, "finished": true, "Ev Idx": 11403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910861.301, "dur": 0.170,
    "args": {
      "Python parent id": 8502, "Python id": 8504, "finished": true, "Ev Idx": 11404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910862.547, "dur": 0.073,
    "args": {
      "Python parent id": 8495, "Python id": 8505, "finished": true, "Ev Idx": 11405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910862.904, "dur": 0.092,
    "args": {
      "Python parent id": 8493, "Python id": 8506, "finished": true, "Ev Idx": 11406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910863.293, "dur": 1.187,
    "args": {
      "Python parent id": 8493, "Python id": 8507, "finished": true, "Ev Idx": 11407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910863.438, "dur": 0.998,
    "args": {
      "Python parent id": 8507, "Python id": 8508, "finished": true, "Ev Idx": 11408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910863.624, "dur": 0.401,
    "args": {
      "Python parent id": 8508, "Python id": 8509, "finished": true, "Ev Idx": 11409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910863.776, "dur": 0.188,
    "args": {
      "Python parent id": 8509, "Python id": 8510, "finished": true, "Ev Idx": 11410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910863.868, "dur": 0.059,
    "args": {
      "Python parent id": 8510, "Python id": 8511, "finished": true, "Ev Idx": 11411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910864.156, "dur": 0.248,
    "args": {
      "Python parent id": 8508, "Python id": 8512, "finished": true, "Ev Idx": 11412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910864.915, "dur": 0.189,
    "args": {
      "Python parent id": 8477, "Python id": 8513, "finished": true, "Ev Idx": 11413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910865.632, "dur": 0.984,
    "args": {
      "Python parent id": 8477, "Python id": 8514, "finished": true, "Ev Idx": 11414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910866.064, "dur": 0.499,
    "args": {
      "Python parent id": 8514, "Python id": 8515, "finished": true, "Ev Idx": 11415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910867.000, "dur": 4.190,
    "args": {
      "Python parent id": 8476, "Python id": 8516, "finished": true, "Ev Idx": 11416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740910872.247, "dur": 91.985,
    "args": {
      "Python parent id": 12, "Python id": 8517, "finished": true, "Ev Idx": 11417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910872.597, "dur": 0.466,
    "args": {
      "Python parent id": 8517, "Python id": 8518, "finished": true, "Ev Idx": 11418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910872.885, "dur": 0.104,
    "args": {
      "Python parent id": 8518, "Python id": 8519, "finished": true, "Ev Idx": 11419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910873.365, "dur": 0.339,
    "args": {
      "Python parent id": 8517, "Python id": 8520, "finished": true, "Ev Idx": 11420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740910873.930, "dur": 0.287,
    "args": {
      "Python parent id": 8517, "Python id": 8521, "finished": true, "Ev Idx": 11421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740910874.468, "dur": 1.557,
    "args": {
      "Python parent id": 8517, "Python id": 8522, "finished": true, "Ev Idx": 11422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740910876.556, "dur": 87.556,
    "args": {
      "Python parent id": 8517, "Python id": 8523, "finished": true, "Ev Idx": 11423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740910877.409, "dur": 86.498,
    "args": {
      "Python parent id": 8523, "Python id": 8524, "finished": true, "Ev Idx": 11424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910877.727, "dur": 86.102,
    "args": {
      "Python parent id": 8524, "Python id": 8525, "finished": true, "Ev Idx": 11425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740910883.073, "dur": 79.294,
    "args": {
      "Python parent id": 8525, "Python id": 8526, "finished": true, "Ev Idx": 11426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740910883.394, "dur": 78.914,
    "args": {
      "Python parent id": 8526, "Python id": 8527, "finished": true, "Ev Idx": 11427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740910883.947, "dur": 78.247,
    "args": {
      "Python parent id": 8527, "Python id": 8528, "finished": true, "Ev Idx": 11428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910884.536, "dur": 0.437,
    "args": {
      "Python parent id": 8528, "Python id": 8529, "finished": true, "Ev Idx": 11429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910884.869, "dur": 0.043,
    "args": {
      "Python parent id": 8529, "Python id": 8530, "finished": true, "Ev Idx": 11430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740910885.274, "dur": 0.404,
    "args": {
      "Python parent id": 8528, "Python id": 8531, "finished": true, "Ev Idx": 11431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740910885.571, "dur": 0.046,
    "args": {
      "Python parent id": 8531, "Python id": 8532, "finished": true, "Ev Idx": 11432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910886.013, "dur": 76.116,
    "args": {
      "Python parent id": 8528, "Python id": 8533, "finished": true, "Ev Idx": 11433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740910886.766, "dur": 74.017,
    "args": {
      "Python parent id": 8533, "Python id": 8534, "finished": true, "Ev Idx": 11434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740910894.005, "dur": 18.432,
    "args": {
      "Python parent id": 8534, "Python id": 8535, "finished": true, "Ev Idx": 11435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740910894.200, "dur": 0.138,
    "args": {
      "Python parent id": 8535, "Python id": 8536, "finished": true, "Ev Idx": 11436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740910894.632, "dur": 17.604,
    "args": {
      "Python parent id": 8535, "Python id": 8537, "finished": true, "Ev Idx": 11437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910894.770, "dur": 7.417,
    "args": {
      "Python parent id": 8537, "Python id": 8538, "finished": true, "Ev Idx": 11438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910920.808, "dur": 7.320,
    "args": {
      "Python parent id": 8534, "Python id": 8539, "finished": true, "Ev Idx": 11439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740910937.129, "dur": 8.583,
    "args": {
      "Python parent id": 8534, "Python id": 8540, "finished": true, "Ev Idx": 11440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740910965.533, "dur": 18.021,
    "args": {
      "Python parent id": 12, "Python id": 8541, "finished": true, "Ev Idx": 11441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740910965.926, "dur": 13.113,
    "args": {
      "Python parent id": 8541, "Python id": 8542, "finished": true, "Ev Idx": 11442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910966.143, "dur": 0.535,
    "args": {
      "Python parent id": 8542, "Python id": 8543, "finished": true, "Ev Idx": 11443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910966.331, "dur": 0.267,
    "args": {
      "Python parent id": 8543, "Python id": 8544, "finished": true, "Ev Idx": 11444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910966.456, "dur": 0.106,
    "args": {
      "Python parent id": 8544, "Python id": 8545, "finished": true, "Ev Idx": 11445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910967.112, "dur": 9.699,
    "args": {
      "Python parent id": 8542, "Python id": 8546, "finished": true, "Ev Idx": 11446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910967.254, "dur": 0.081,
    "args": {
      "Python parent id": 8546, "Python id": 8547, "finished": true, "Ev Idx": 11447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910967.473, "dur": 0.065,
    "args": {
      "Python parent id": 8546, "Python id": 8548, "finished": true, "Ev Idx": 11448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910967.717, "dur": 0.060,
    "args": {
      "Python parent id": 8546, "Python id": 8549, "finished": true, "Ev Idx": 11449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910968.350, "dur": 0.065,
    "args": {
      "Python parent id": 8546, "Python id": 8550, "finished": true, "Ev Idx": 11450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910968.725, "dur": 0.072,
    "args": {
      "Python parent id": 8546, "Python id": 8551, "finished": true, "Ev Idx": 11451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910969.192, "dur": 7.559,
    "args": {
      "Python parent id": 8546, "Python id": 8552, "finished": true, "Ev Idx": 11452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910969.300, "dur": 0.064,
    "args": {
      "Python parent id": 8552, "Python id": 8553, "finished": true, "Ev Idx": 11453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910969.534, "dur": 0.056,
    "args": {
      "Python parent id": 8552, "Python id": 8554, "finished": true, "Ev Idx": 11454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910969.783, "dur": 0.059,
    "args": {
      "Python parent id": 8552, "Python id": 8555, "finished": true, "Ev Idx": 11455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740910970.149, "dur": 0.035,
    "args": {
      "Python parent id": 8552, "Python id": 8556, "finished": true, "Ev Idx": 11456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740910970.510, "dur": 6.144,
    "args": {
      "Python parent id": 8552, "Python id": 8557, "finished": true, "Ev Idx": 11457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740910970.827, "dur": 5.743,
    "args": {
      "Python parent id": 8557, "Python id": 8558, "finished": true, "Ev Idx": 11458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740910970.970, "dur": 3.656,
    "args": {
      "Python parent id": 8558, "Python id": 8559, "finished": true, "Ev Idx": 11459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740910971.157, "dur": 3.381,
    "args": {
      "Python parent id": 8559, "Python id": 8560, "finished": true, "Ev Idx": 11460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740910971.306, "dur": 0.280,
    "args": {
      "Python parent id": 8560, "Python id": 8561, "finished": true, "Ev Idx": 11461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740910971.446, "dur": 0.071,
    "args": {
      "Python parent id": 8561, "Python id": 8562, "finished": true, "Ev Idx": 11462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740910971.752, "dur": 2.365,
    "args": {
      "Python parent id": 8560, "Python id": 8563, "finished": true, "Ev Idx": 11463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740910971.983, "dur": 2.052,
    "args": {
      "Python parent id": 8563, "Python id": 8564, "finished": true, "Ev Idx": 11464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740910972.214, "dur": 1.769,
    "args": {
      "Python parent id": 8564, "Python id": 8565, "finished": true, "Ev Idx": 11465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740910972.365, "dur": 1.281,
    "args": {
      "Python parent id": 8565, "Python id": 8566, "finished": true, "Ev Idx": 11466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740910972.569, "dur": 0.537,
    "args": {
      "Python parent id": 8566, "Python id": 8567, "finished": true, "Ev Idx": 11467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910972.659, "dur": 0.042,
    "args": {
      "Python parent id": 8567, "Python id": 8568, "finished": true, "Ev Idx": 11468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740910972.875, "dur": 0.150,
    "args": {
      "Python parent id": 8567, "Python id": 8569, "finished": true, "Ev Idx": 11469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740910974.323, "dur": 0.160,
    "args": {
      "Python parent id": 8560, "Python id": 8570, "finished": true, "Ev Idx": 11470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740910974.809, "dur": 0.134,
    "args": {
      "Python parent id": 8558, "Python id": 8571, "finished": true, "Ev Idx": 11471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740910975.270, "dur": 1.235,
    "args": {
      "Python parent id": 8558, "Python id": 8572, "finished": true, "Ev Idx": 11472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740910975.427, "dur": 1.011,
    "args": {
      "Python parent id": 8572, "Python id": 8573, "finished": true, "Ev Idx": 11473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740910975.570, "dur": 0.409,
    "args": {
      "Python parent id": 8573, "Python id": 8574, "finished": true, "Ev Idx": 11474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740910975.706, "dur": 0.186,
    "args": {
      "Python parent id": 8574, "Python id": 8575, "finished": true, "Ev Idx": 11475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740910975.801, "dur": 0.057,
    "args": {
      "Python parent id": 8575, "Python id": 8576, "finished": true, "Ev Idx": 11476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740910976.115, "dur": 0.291,
    "args": {
      "Python parent id": 8573, "Python id": 8577, "finished": true, "Ev Idx": 11477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740910976.946, "dur": 0.211,
    "args": {
      "Python parent id": 8542, "Python id": 8578, "finished": true, "Ev Idx": 11478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740910977.808, "dur": 1.062,
    "args": {
      "Python parent id": 8542, "Python id": 8579, "finished": true, "Ev Idx": 11479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740910978.270, "dur": 0.549,
    "args": {
      "Python parent id": 8579, "Python id": 8580, "finished": true, "Ev Idx": 11480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740910979.373, "dur": 4.060,
    "args": {
      "Python parent id": 8541, "Python id": 8581, "finished": true, "Ev Idx": 11481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740910984.187, "dur": 7.974,
    "args": {
      "Python parent id": 12, "Python id": 8582, "finished": true, "Ev Idx": 11482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740910992.642, "dur": 0.116,
    "args": {
      "Python parent id": 12, "Python id": 8583, "finished": true, "Ev Idx": 11483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910993.482, "dur": 0.185,
    "args": {
      "Python parent id": 12, "Python id": 8584, "finished": true, "Ev Idx": 11484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740910994.191, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 8585, "finished": true, "Ev Idx": 11485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740910994.663, "dur": 10.120,
    "args": {
      "Python parent id": 12, "Python id": 8586, "finished": true, "Ev Idx": 11486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911006.154, "dur": 16.352,
    "args": {
      "Python parent id": 12, "Python id": 8587, "finished": true, "Ev Idx": 11487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911006.609, "dur": 11.411,
    "args": {
      "Python parent id": 8587, "Python id": 8588, "finished": true, "Ev Idx": 11488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911006.754, "dur": 0.458,
    "args": {
      "Python parent id": 8588, "Python id": 8589, "finished": true, "Ev Idx": 11489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911006.901, "dur": 0.233,
    "args": {
      "Python parent id": 8589, "Python id": 8590, "finished": true, "Ev Idx": 11490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911007.001, "dur": 0.096,
    "args": {
      "Python parent id": 8590, "Python id": 8591, "finished": true, "Ev Idx": 11491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911007.514, "dur": 8.539,
    "args": {
      "Python parent id": 8588, "Python id": 8592, "finished": true, "Ev Idx": 11492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911007.599, "dur": 0.074,
    "args": {
      "Python parent id": 8592, "Python id": 8593, "finished": true, "Ev Idx": 11493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911007.790, "dur": 0.062,
    "args": {
      "Python parent id": 8592, "Python id": 8594, "finished": true, "Ev Idx": 11494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911007.997, "dur": 0.063,
    "args": {
      "Python parent id": 8592, "Python id": 8595, "finished": true, "Ev Idx": 11495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911008.427, "dur": 0.036,
    "args": {
      "Python parent id": 8592, "Python id": 8596, "finished": true, "Ev Idx": 11496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911008.751, "dur": 0.083,
    "args": {
      "Python parent id": 8592, "Python id": 8597, "finished": true, "Ev Idx": 11497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911009.227, "dur": 6.754,
    "args": {
      "Python parent id": 8592, "Python id": 8598, "finished": true, "Ev Idx": 11498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911009.339, "dur": 0.062,
    "args": {
      "Python parent id": 8598, "Python id": 8599, "finished": true, "Ev Idx": 11499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911009.562, "dur": 0.062,
    "args": {
      "Python parent id": 8598, "Python id": 8600, "finished": true, "Ev Idx": 11500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911009.802, "dur": 0.070,
    "args": {
      "Python parent id": 8598, "Python id": 8601, "finished": true, "Ev Idx": 11501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911010.175, "dur": 0.042,
    "args": {
      "Python parent id": 8598, "Python id": 8602, "finished": true, "Ev Idx": 11502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911010.579, "dur": 5.316,
    "args": {
      "Python parent id": 8598, "Python id": 8603, "finished": true, "Ev Idx": 11503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911010.821, "dur": 5.015,
    "args": {
      "Python parent id": 8603, "Python id": 8604, "finished": true, "Ev Idx": 11504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911010.952, "dur": 3.078,
    "args": {
      "Python parent id": 8604, "Python id": 8605, "finished": true, "Ev Idx": 11505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911011.163, "dur": 2.820,
    "args": {
      "Python parent id": 8605, "Python id": 8606, "finished": true, "Ev Idx": 11506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911011.336, "dur": 0.244,
    "args": {
      "Python parent id": 8606, "Python id": 8607, "finished": true, "Ev Idx": 11507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911011.444, "dur": 0.075,
    "args": {
      "Python parent id": 8607, "Python id": 8608, "finished": true, "Ev Idx": 11508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911011.754, "dur": 1.924,
    "args": {
      "Python parent id": 8606, "Python id": 8609, "finished": true, "Ev Idx": 11509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911011.938, "dur": 1.676,
    "args": {
      "Python parent id": 8609, "Python id": 8610, "finished": true, "Ev Idx": 11510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911012.133, "dur": 1.432,
    "args": {
      "Python parent id": 8610, "Python id": 8611, "finished": true, "Ev Idx": 11511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911012.253, "dur": 1.106,
    "args": {
      "Python parent id": 8611, "Python id": 8612, "finished": true, "Ev Idx": 11512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911012.426, "dur": 0.475,
    "args": {
      "Python parent id": 8612, "Python id": 8613, "finished": true, "Ev Idx": 11513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911012.513, "dur": 0.052,
    "args": {
      "Python parent id": 8613, "Python id": 8614, "finished": true, "Ev Idx": 11514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911012.723, "dur": 0.128,
    "args": {
      "Python parent id": 8613, "Python id": 8615, "finished": true, "Ev Idx": 11515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911013.874, "dur": 0.075,
    "args": {
      "Python parent id": 8606, "Python id": 8616, "finished": true, "Ev Idx": 11516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911014.187, "dur": 0.113,
    "args": {
      "Python parent id": 8604, "Python id": 8617, "finished": true, "Ev Idx": 11517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911014.596, "dur": 1.191,
    "args": {
      "Python parent id": 8604, "Python id": 8618, "finished": true, "Ev Idx": 11518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911014.750, "dur": 0.992,
    "args": {
      "Python parent id": 8618, "Python id": 8619, "finished": true, "Ev Idx": 11519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911014.891, "dur": 0.453,
    "args": {
      "Python parent id": 8619, "Python id": 8620, "finished": true, "Ev Idx": 11520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911015.067, "dur": 0.207,
    "args": {
      "Python parent id": 8620, "Python id": 8621, "finished": true, "Ev Idx": 11521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911015.169, "dur": 0.069,
    "args": {
      "Python parent id": 8621, "Python id": 8622, "finished": true, "Ev Idx": 11522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911015.477, "dur": 0.232,
    "args": {
      "Python parent id": 8619, "Python id": 8623, "finished": true, "Ev Idx": 11523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911016.203, "dur": 0.140,
    "args": {
      "Python parent id": 8588, "Python id": 8624, "finished": true, "Ev Idx": 11524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911016.879, "dur": 1.002,
    "args": {
      "Python parent id": 8588, "Python id": 8625, "finished": true, "Ev Idx": 11525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911017.351, "dur": 0.465,
    "args": {
      "Python parent id": 8625, "Python id": 8626, "finished": true, "Ev Idx": 11526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911018.310, "dur": 4.139,
    "args": {
      "Python parent id": 8587, "Python id": 8627, "finished": true, "Ev Idx": 11527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911023.499, "dur": 90.158,
    "args": {
      "Python parent id": 12, "Python id": 8628, "finished": true, "Ev Idx": 11528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911023.904, "dur": 0.453,
    "args": {
      "Python parent id": 8628, "Python id": 8629, "finished": true, "Ev Idx": 11529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911024.189, "dur": 0.094,
    "args": {
      "Python parent id": 8629, "Python id": 8630, "finished": true, "Ev Idx": 11530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911024.639, "dur": 0.374,
    "args": {
      "Python parent id": 8628, "Python id": 8631, "finished": true, "Ev Idx": 11531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911025.224, "dur": 0.267,
    "args": {
      "Python parent id": 8628, "Python id": 8632, "finished": true, "Ev Idx": 11532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911025.738, "dur": 1.571,
    "args": {
      "Python parent id": 8628, "Python id": 8633, "finished": true, "Ev Idx": 11533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911027.925, "dur": 85.660,
    "args": {
      "Python parent id": 8628, "Python id": 8634, "finished": true, "Ev Idx": 11534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911028.811, "dur": 84.556,
    "args": {
      "Python parent id": 8634, "Python id": 8635, "finished": true, "Ev Idx": 11535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911029.106, "dur": 84.176,
    "args": {
      "Python parent id": 8635, "Python id": 8636, "finished": true, "Ev Idx": 11536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911034.188, "dur": 77.675,
    "args": {
      "Python parent id": 8636, "Python id": 8637, "finished": true, "Ev Idx": 11537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911034.522, "dur": 77.258,
    "args": {
      "Python parent id": 8637, "Python id": 8638, "finished": true, "Ev Idx": 11538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911035.083, "dur": 76.574,
    "args": {
      "Python parent id": 8638, "Python id": 8639, "finished": true, "Ev Idx": 11539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911035.647, "dur": 0.378,
    "args": {
      "Python parent id": 8639, "Python id": 8640, "finished": true, "Ev Idx": 11540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911035.914, "dur": 0.047,
    "args": {
      "Python parent id": 8640, "Python id": 8641, "finished": true, "Ev Idx": 11541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911036.349, "dur": 0.363,
    "args": {
      "Python parent id": 8639, "Python id": 8642, "finished": true, "Ev Idx": 11542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911036.609, "dur": 0.045,
    "args": {
      "Python parent id": 8642, "Python id": 8643, "finished": true, "Ev Idx": 11543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911037.042, "dur": 74.512,
    "args": {
      "Python parent id": 8639, "Python id": 8644, "finished": true, "Ev Idx": 11544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911037.787, "dur": 72.413,
    "args": {
      "Python parent id": 8644, "Python id": 8645, "finished": true, "Ev Idx": 11545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911044.958, "dur": 18.524,
    "args": {
      "Python parent id": 8645, "Python id": 8646, "finished": true, "Ev Idx": 11546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911045.118, "dur": 0.142,
    "args": {
      "Python parent id": 8646, "Python id": 8647, "finished": true, "Ev Idx": 11547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911045.529, "dur": 17.765,
    "args": {
      "Python parent id": 8646, "Python id": 8648, "finished": true, "Ev Idx": 11548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911045.660, "dur": 7.658,
    "args": {
      "Python parent id": 8648, "Python id": 8649, "finished": true, "Ev Idx": 11549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911071.395, "dur": 7.140,
    "args": {
      "Python parent id": 8645, "Python id": 8650, "finished": true, "Ev Idx": 11550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911087.331, "dur": 8.277,
    "args": {
      "Python parent id": 8645, "Python id": 8651, "finished": true, "Ev Idx": 11551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911114.870, "dur": 17.620,
    "args": {
      "Python parent id": 12, "Python id": 8652, "finished": true, "Ev Idx": 11552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911115.294, "dur": 13.107,
    "args": {
      "Python parent id": 8652, "Python id": 8653, "finished": true, "Ev Idx": 11553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911115.499, "dur": 0.511,
    "args": {
      "Python parent id": 8653, "Python id": 8654, "finished": true, "Ev Idx": 11554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911115.691, "dur": 0.226,
    "args": {
      "Python parent id": 8654, "Python id": 8655, "finished": true, "Ev Idx": 11555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911115.809, "dur": 0.072,
    "args": {
      "Python parent id": 8655, "Python id": 8656, "finished": true, "Ev Idx": 11556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911116.360, "dur": 9.715,
    "args": {
      "Python parent id": 8653, "Python id": 8657, "finished": true, "Ev Idx": 11557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911116.451, "dur": 0.064,
    "args": {
      "Python parent id": 8657, "Python id": 8658, "finished": true, "Ev Idx": 11558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911116.652, "dur": 0.063,
    "args": {
      "Python parent id": 8657, "Python id": 8659, "finished": true, "Ev Idx": 11559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911116.865, "dur": 0.067,
    "args": {
      "Python parent id": 8657, "Python id": 8660, "finished": true, "Ev Idx": 11560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911117.439, "dur": 0.054,
    "args": {
      "Python parent id": 8657, "Python id": 8661, "finished": true, "Ev Idx": 11561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911117.805, "dur": 0.078,
    "args": {
      "Python parent id": 8657, "Python id": 8662, "finished": true, "Ev Idx": 11562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911118.301, "dur": 7.702,
    "args": {
      "Python parent id": 8657, "Python id": 8663, "finished": true, "Ev Idx": 11563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911118.421, "dur": 0.081,
    "args": {
      "Python parent id": 8663, "Python id": 8664, "finished": true, "Ev Idx": 11564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911118.679, "dur": 0.064,
    "args": {
      "Python parent id": 8663, "Python id": 8665, "finished": true, "Ev Idx": 11565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911118.936, "dur": 0.075,
    "args": {
      "Python parent id": 8663, "Python id": 8666, "finished": true, "Ev Idx": 11566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911119.430, "dur": 0.038,
    "args": {
      "Python parent id": 8663, "Python id": 8667, "finished": true, "Ev Idx": 11567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911119.848, "dur": 6.069,
    "args": {
      "Python parent id": 8663, "Python id": 8668, "finished": true, "Ev Idx": 11568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911120.141, "dur": 5.685,
    "args": {
      "Python parent id": 8668, "Python id": 8669, "finished": true, "Ev Idx": 11569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911120.327, "dur": 3.640,
    "args": {
      "Python parent id": 8669, "Python id": 8670, "finished": true, "Ev Idx": 11570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911120.544, "dur": 3.344,
    "args": {
      "Python parent id": 8670, "Python id": 8671, "finished": true, "Ev Idx": 11571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911120.679, "dur": 0.305,
    "args": {
      "Python parent id": 8671, "Python id": 8672, "finished": true, "Ev Idx": 11572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911120.825, "dur": 0.078,
    "args": {
      "Python parent id": 8672, "Python id": 8673, "finished": true, "Ev Idx": 11573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911121.173, "dur": 2.309,
    "args": {
      "Python parent id": 8671, "Python id": 8674, "finished": true, "Ev Idx": 11574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911121.421, "dur": 1.960,
    "args": {
      "Python parent id": 8674, "Python id": 8675, "finished": true, "Ev Idx": 11575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911121.655, "dur": 1.659,
    "args": {
      "Python parent id": 8675, "Python id": 8676, "finished": true, "Ev Idx": 11576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911121.805, "dur": 1.250,
    "args": {
      "Python parent id": 8676, "Python id": 8677, "finished": true, "Ev Idx": 11577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911122.044, "dur": 0.501,
    "args": {
      "Python parent id": 8677, "Python id": 8678, "finished": true, "Ev Idx": 11578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911122.145, "dur": 0.042,
    "args": {
      "Python parent id": 8678, "Python id": 8679, "finished": true, "Ev Idx": 11579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911122.372, "dur": 0.127,
    "args": {
      "Python parent id": 8678, "Python id": 8680, "finished": true, "Ev Idx": 11580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911123.659, "dur": 0.168,
    "args": {
      "Python parent id": 8671, "Python id": 8681, "finished": true, "Ev Idx": 11581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911124.126, "dur": 0.113,
    "args": {
      "Python parent id": 8669, "Python id": 8682, "finished": true, "Ev Idx": 11582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911124.577, "dur": 1.177,
    "args": {
      "Python parent id": 8669, "Python id": 8683, "finished": true, "Ev Idx": 11583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911124.723, "dur": 0.994,
    "args": {
      "Python parent id": 8683, "Python id": 8684, "finished": true, "Ev Idx": 11584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911124.867, "dur": 0.402,
    "args": {
      "Python parent id": 8684, "Python id": 8685, "finished": true, "Ev Idx": 11585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911125.019, "dur": 0.188,
    "args": {
      "Python parent id": 8685, "Python id": 8686, "finished": true, "Ev Idx": 11586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911125.117, "dur": 0.055,
    "args": {
      "Python parent id": 8686, "Python id": 8687, "finished": true, "Ev Idx": 11587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911125.400, "dur": 0.264,
    "args": {
      "Python parent id": 8684, "Python id": 8688, "finished": true, "Ev Idx": 11588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911126.208, "dur": 0.262,
    "args": {
      "Python parent id": 8653, "Python id": 8689, "finished": true, "Ev Idx": 11589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911127.140, "dur": 1.068,
    "args": {
      "Python parent id": 8653, "Python id": 8690, "finished": true, "Ev Idx": 11590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911127.633, "dur": 0.524,
    "args": {
      "Python parent id": 8690, "Python id": 8691, "finished": true, "Ev Idx": 11591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911128.706, "dur": 3.733,
    "args": {
      "Python parent id": 8652, "Python id": 8692, "finished": true, "Ev Idx": 11592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911133.120, "dur": 7.552,
    "args": {
      "Python parent id": 12, "Python id": 8693, "finished": true, "Ev Idx": 11593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911141.205, "dur": 0.118,
    "args": {
      "Python parent id": 12, "Python id": 8694, "finished": true, "Ev Idx": 11594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911141.993, "dur": 0.219,
    "args": {
      "Python parent id": 12, "Python id": 8695, "finished": true, "Ev Idx": 11595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911142.724, "dur": 0.069,
    "args": {
      "Python parent id": 12, "Python id": 8696, "finished": true, "Ev Idx": 11596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911143.208, "dur": 9.793,
    "args": {
      "Python parent id": 12, "Python id": 8697, "finished": true, "Ev Idx": 11597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911154.300, "dur": 16.127,
    "args": {
      "Python parent id": 12, "Python id": 8698, "finished": true, "Ev Idx": 11598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911154.669, "dur": 12.183,
    "args": {
      "Python parent id": 8698, "Python id": 8699, "finished": true, "Ev Idx": 11599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911154.832, "dur": 0.514,
    "args": {
      "Python parent id": 8699, "Python id": 8700, "finished": true, "Ev Idx": 11600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911154.989, "dur": 0.277,
    "args": {
      "Python parent id": 8700, "Python id": 8701, "finished": true, "Ev Idx": 11601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911155.107, "dur": 0.126,
    "args": {
      "Python parent id": 8701, "Python id": 8702, "finished": true, "Ev Idx": 11602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911155.627, "dur": 9.222,
    "args": {
      "Python parent id": 8699, "Python id": 8703, "finished": true, "Ev Idx": 11603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911155.749, "dur": 0.066,
    "args": {
      "Python parent id": 8703, "Python id": 8704, "finished": true, "Ev Idx": 11604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911155.961, "dur": 0.060,
    "args": {
      "Python parent id": 8703, "Python id": 8705, "finished": true, "Ev Idx": 11605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911156.200, "dur": 0.065,
    "args": {
      "Python parent id": 8703, "Python id": 8706, "finished": true, "Ev Idx": 11606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911156.680, "dur": 0.042,
    "args": {
      "Python parent id": 8703, "Python id": 8707, "finished": true, "Ev Idx": 11607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911157.015, "dur": 0.081,
    "args": {
      "Python parent id": 8703, "Python id": 8708, "finished": true, "Ev Idx": 11608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911157.527, "dur": 7.255,
    "args": {
      "Python parent id": 8703, "Python id": 8709, "finished": true, "Ev Idx": 11609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911157.623, "dur": 0.062,
    "args": {
      "Python parent id": 8709, "Python id": 8710, "finished": true, "Ev Idx": 11610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911157.826, "dur": 0.061,
    "args": {
      "Python parent id": 8709, "Python id": 8711, "finished": true, "Ev Idx": 11611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911158.073, "dur": 0.068,
    "args": {
      "Python parent id": 8709, "Python id": 8712, "finished": true, "Ev Idx": 11612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911158.513, "dur": 0.055,
    "args": {
      "Python parent id": 8709, "Python id": 8713, "finished": true, "Ev Idx": 11613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911158.931, "dur": 5.788,
    "args": {
      "Python parent id": 8709, "Python id": 8714, "finished": true, "Ev Idx": 11614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911159.182, "dur": 5.474,
    "args": {
      "Python parent id": 8714, "Python id": 8715, "finished": true, "Ev Idx": 11615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911159.307, "dur": 3.678,
    "args": {
      "Python parent id": 8715, "Python id": 8716, "finished": true, "Ev Idx": 11616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911159.515, "dur": 3.427,
    "args": {
      "Python parent id": 8716, "Python id": 8717, "finished": true, "Ev Idx": 11617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911159.680, "dur": 0.297,
    "args": {
      "Python parent id": 8717, "Python id": 8718, "finished": true, "Ev Idx": 11618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911159.835, "dur": 0.070,
    "args": {
      "Python parent id": 8718, "Python id": 8719, "finished": true, "Ev Idx": 11619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911160.124, "dur": 2.415,
    "args": {
      "Python parent id": 8717, "Python id": 8720, "finished": true, "Ev Idx": 11620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911160.293, "dur": 2.158,
    "args": {
      "Python parent id": 8720, "Python id": 8721, "finished": true, "Ev Idx": 11621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911160.493, "dur": 1.894,
    "args": {
      "Python parent id": 8721, "Python id": 8722, "finished": true, "Ev Idx": 11622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911160.968, "dur": 1.218,
    "args": {
      "Python parent id": 8722, "Python id": 8723, "finished": true, "Ev Idx": 11623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911161.180, "dur": 0.494,
    "args": {
      "Python parent id": 8723, "Python id": 8724, "finished": true, "Ev Idx": 11624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911161.275, "dur": 0.044,
    "args": {
      "Python parent id": 8724, "Python id": 8725, "finished": true, "Ev Idx": 11625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911161.477, "dur": 0.151,
    "args": {
      "Python parent id": 8724, "Python id": 8726, "finished": true, "Ev Idx": 11626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911162.780, "dur": 0.109,
    "args": {
      "Python parent id": 8717, "Python id": 8727, "finished": true, "Ev Idx": 11627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911163.168, "dur": 0.076,
    "args": {
      "Python parent id": 8715, "Python id": 8728, "finished": true, "Ev Idx": 11628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911163.488, "dur": 1.124,
    "args": {
      "Python parent id": 8715, "Python id": 8729, "finished": true, "Ev Idx": 11629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911163.654, "dur": 0.911,
    "args": {
      "Python parent id": 8729, "Python id": 8730, "finished": true, "Ev Idx": 11630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911163.809, "dur": 0.382,
    "args": {
      "Python parent id": 8730, "Python id": 8731, "finished": true, "Ev Idx": 11631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911163.950, "dur": 0.182,
    "args": {
      "Python parent id": 8731, "Python id": 8732, "finished": true, "Ev Idx": 11632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911164.043, "dur": 0.058,
    "args": {
      "Python parent id": 8732, "Python id": 8733, "finished": true, "Ev Idx": 11633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911164.326, "dur": 0.211,
    "args": {
      "Python parent id": 8730, "Python id": 8734, "finished": true, "Ev Idx": 11634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911164.980, "dur": 0.177,
    "args": {
      "Python parent id": 8699, "Python id": 8735, "finished": true, "Ev Idx": 11635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911165.725, "dur": 1.002,
    "args": {
      "Python parent id": 8699, "Python id": 8736, "finished": true, "Ev Idx": 11636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911166.175, "dur": 0.493,
    "args": {
      "Python parent id": 8736, "Python id": 8737, "finished": true, "Ev Idx": 11637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911167.101, "dur": 3.274,
    "args": {
      "Python parent id": 8698, "Python id": 8738, "finished": true, "Ev Idx": 11638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911171.336, "dur": 113.859,
    "args": {
      "Python parent id": 12, "Python id": 8739, "finished": true, "Ev Idx": 11639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911171.715, "dur": 0.433,
    "args": {
      "Python parent id": 8739, "Python id": 8740, "finished": true, "Ev Idx": 11640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911171.992, "dur": 0.077,
    "args": {
      "Python parent id": 8740, "Python id": 8741, "finished": true, "Ev Idx": 11641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911172.430, "dur": 0.357,
    "args": {
      "Python parent id": 8739, "Python id": 8742, "finished": true, "Ev Idx": 11642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911173.020, "dur": 0.247,
    "args": {
      "Python parent id": 8739, "Python id": 8743, "finished": true, "Ev Idx": 11643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911173.477, "dur": 1.437,
    "args": {
      "Python parent id": 8739, "Python id": 8744, "finished": true, "Ev Idx": 11644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911175.550, "dur": 109.553,
    "args": {
      "Python parent id": 8739, "Python id": 8745, "finished": true, "Ev Idx": 11645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911176.381, "dur": 108.479,
    "args": {
      "Python parent id": 8745, "Python id": 8746, "finished": true, "Ev Idx": 11646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911176.717, "dur": 108.085,
    "args": {
      "Python parent id": 8746, "Python id": 8747, "finished": true, "Ev Idx": 11647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911181.752, "dur": 101.371,
    "args": {
      "Python parent id": 8747, "Python id": 8748, "finished": true, "Ev Idx": 11648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911182.073, "dur": 100.924,
    "args": {
      "Python parent id": 8748, "Python id": 8749, "finished": true, "Ev Idx": 11649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911182.607, "dur": 100.321,
    "args": {
      "Python parent id": 8749, "Python id": 8750, "finished": true, "Ev Idx": 11650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911183.084, "dur": 0.388,
    "args": {
      "Python parent id": 8750, "Python id": 8751, "finished": true, "Ev Idx": 11651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911183.353, "dur": 0.048,
    "args": {
      "Python parent id": 8751, "Python id": 8752, "finished": true, "Ev Idx": 11652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911183.807, "dur": 0.331,
    "args": {
      "Python parent id": 8750, "Python id": 8753, "finished": true, "Ev Idx": 11653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911184.043, "dur": 0.038,
    "args": {
      "Python parent id": 8753, "Python id": 8754, "finished": true, "Ev Idx": 11654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911184.510, "dur": 98.243,
    "args": {
      "Python parent id": 8750, "Python id": 8755, "finished": true, "Ev Idx": 11655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911185.197, "dur": 96.170,
    "args": {
      "Python parent id": 8755, "Python id": 8756, "finished": true, "Ev Idx": 11656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911208.365, "dur": 20.188,
    "args": {
      "Python parent id": 8756, "Python id": 8757, "finished": true, "Ev Idx": 11657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911208.554, "dur": 0.158,
    "args": {
      "Python parent id": 8757, "Python id": 8758, "finished": true, "Ev Idx": 11658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911209.062, "dur": 19.312,
    "args": {
      "Python parent id": 8757, "Python id": 8759, "finished": true, "Ev Idx": 11659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911209.167, "dur": 8.529,
    "args": {
      "Python parent id": 8759, "Python id": 8760, "finished": true, "Ev Idx": 11660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911238.195, "dur": 8.129,
    "args": {
      "Python parent id": 8756, "Python id": 8761, "finished": true, "Ev Idx": 11661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911256.593, "dur": 8.907,
    "args": {
      "Python parent id": 8756, "Python id": 8762, "finished": true, "Ev Idx": 11662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911286.735, "dur": 19.682,
    "args": {
      "Python parent id": 12, "Python id": 8763, "finished": true, "Ev Idx": 11663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911287.095, "dur": 15.006,
    "args": {
      "Python parent id": 8763, "Python id": 8764, "finished": true, "Ev Idx": 11664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911287.332, "dur": 0.553,
    "args": {
      "Python parent id": 8764, "Python id": 8765, "finished": true, "Ev Idx": 11665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911287.490, "dur": 0.298,
    "args": {
      "Python parent id": 8765, "Python id": 8766, "finished": true, "Ev Idx": 11666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911287.631, "dur": 0.120,
    "args": {
      "Python parent id": 8766, "Python id": 8767, "finished": true, "Ev Idx": 11667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911288.308, "dur": 11.050,
    "args": {
      "Python parent id": 8764, "Python id": 8768, "finished": true, "Ev Idx": 11668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911288.449, "dur": 0.092,
    "args": {
      "Python parent id": 8768, "Python id": 8769, "finished": true, "Ev Idx": 11669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911288.700, "dur": 0.066,
    "args": {
      "Python parent id": 8768, "Python id": 8770, "finished": true, "Ev Idx": 11670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911288.939, "dur": 0.063,
    "args": {
      "Python parent id": 8768, "Python id": 8771, "finished": true, "Ev Idx": 11671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911289.479, "dur": 0.050,
    "args": {
      "Python parent id": 8768, "Python id": 8772, "finished": true, "Ev Idx": 11672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911289.822, "dur": 0.070,
    "args": {
      "Python parent id": 8768, "Python id": 8773, "finished": true, "Ev Idx": 11673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911290.287, "dur": 8.979,
    "args": {
      "Python parent id": 8768, "Python id": 8774, "finished": true, "Ev Idx": 11674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911290.378, "dur": 0.062,
    "args": {
      "Python parent id": 8774, "Python id": 8775, "finished": true, "Ev Idx": 11675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911290.619, "dur": 0.054,
    "args": {
      "Python parent id": 8774, "Python id": 8776, "finished": true, "Ev Idx": 11676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911290.890, "dur": 0.065,
    "args": {
      "Python parent id": 8774, "Python id": 8777, "finished": true, "Ev Idx": 11677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911291.403, "dur": 0.054,
    "args": {
      "Python parent id": 8774, "Python id": 8778, "finished": true, "Ev Idx": 11678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911291.798, "dur": 7.415,
    "args": {
      "Python parent id": 8774, "Python id": 8779, "finished": true, "Ev Idx": 11679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911292.127, "dur": 6.978,
    "args": {
      "Python parent id": 8779, "Python id": 8780, "finished": true, "Ev Idx": 11680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911292.289, "dur": 4.268,
    "args": {
      "Python parent id": 8780, "Python id": 8781, "finished": true, "Ev Idx": 11681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911292.493, "dur": 4.003,
    "args": {
      "Python parent id": 8781, "Python id": 8782, "finished": true, "Ev Idx": 11682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911292.682, "dur": 0.291,
    "args": {
      "Python parent id": 8782, "Python id": 8783, "finished": true, "Ev Idx": 11683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911292.809, "dur": 0.086,
    "args": {
      "Python parent id": 8783, "Python id": 8784, "finished": true, "Ev Idx": 11684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911293.148, "dur": 2.902,
    "args": {
      "Python parent id": 8782, "Python id": 8785, "finished": true, "Ev Idx": 11685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911293.425, "dur": 2.532,
    "args": {
      "Python parent id": 8785, "Python id": 8786, "finished": true, "Ev Idx": 11686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911293.765, "dur": 2.129,
    "args": {
      "Python parent id": 8786, "Python id": 8787, "finished": true, "Ev Idx": 11687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911293.944, "dur": 1.611,
    "args": {
      "Python parent id": 8787, "Python id": 8788, "finished": true, "Ev Idx": 11688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911294.238, "dur": 0.632,
    "args": {
      "Python parent id": 8788, "Python id": 8789, "finished": true, "Ev Idx": 11689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911294.326, "dur": 0.057,
    "args": {
      "Python parent id": 8789, "Python id": 8790, "finished": true, "Ev Idx": 11690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911294.596, "dur": 0.207,
    "args": {
      "Python parent id": 8789, "Python id": 8791, "finished": true, "Ev Idx": 11691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911296.246, "dur": 0.175,
    "args": {
      "Python parent id": 8782, "Python id": 8792, "finished": true, "Ev Idx": 11692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911297.109, "dur": 0.116,
    "args": {
      "Python parent id": 8780, "Python id": 8793, "finished": true, "Ev Idx": 11693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911297.569, "dur": 1.459,
    "args": {
      "Python parent id": 8780, "Python id": 8794, "finished": true, "Ev Idx": 11694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911297.751, "dur": 1.211,
    "args": {
      "Python parent id": 8794, "Python id": 8795, "finished": true, "Ev Idx": 11695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911297.935, "dur": 0.462,
    "args": {
      "Python parent id": 8795, "Python id": 8796, "finished": true, "Ev Idx": 11696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911298.089, "dur": 0.222,
    "args": {
      "Python parent id": 8796, "Python id": 8797, "finished": true, "Ev Idx": 11697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911298.201, "dur": 0.078,
    "args": {
      "Python parent id": 8797, "Python id": 8798, "finished": true, "Ev Idx": 11698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911298.566, "dur": 0.343,
    "args": {
      "Python parent id": 8795, "Python id": 8799, "finished": true, "Ev Idx": 11699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911299.518, "dur": 0.279,
    "args": {
      "Python parent id": 8764, "Python id": 8800, "finished": true, "Ev Idx": 11700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911300.586, "dur": 1.347,
    "args": {
      "Python parent id": 8764, "Python id": 8801, "finished": true, "Ev Idx": 11701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911301.176, "dur": 0.693,
    "args": {
      "Python parent id": 8801, "Python id": 8802, "finished": true, "Ev Idx": 11702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911302.534, "dur": 3.774,
    "args": {
      "Python parent id": 8763, "Python id": 8803, "finished": true, "Ev Idx": 11703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911307.201, "dur": 7.288,
    "args": {
      "Python parent id": 12, "Python id": 8804, "finished": true, "Ev Idx": 11704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911314.983, "dur": 0.129,
    "args": {
      "Python parent id": 12, "Python id": 8805, "finished": true, "Ev Idx": 11705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911315.761, "dur": 0.230,
    "args": {
      "Python parent id": 12, "Python id": 8806, "finished": true, "Ev Idx": 11706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911316.515, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 8807, "finished": true, "Ev Idx": 11707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911317.013, "dur": 10.650,
    "args": {
      "Python parent id": 12, "Python id": 8808, "finished": true, "Ev Idx": 11708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911329.111, "dur": 17.160,
    "args": {
      "Python parent id": 12, "Python id": 8809, "finished": true, "Ev Idx": 11709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911329.484, "dur": 12.057,
    "args": {
      "Python parent id": 8809, "Python id": 8810, "finished": true, "Ev Idx": 11710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911329.663, "dur": 0.505,
    "args": {
      "Python parent id": 8810, "Python id": 8811, "finished": true, "Ev Idx": 11711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911329.863, "dur": 0.228,
    "args": {
      "Python parent id": 8811, "Python id": 8812, "finished": true, "Ev Idx": 11712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911329.963, "dur": 0.094,
    "args": {
      "Python parent id": 8812, "Python id": 8813, "finished": true, "Ev Idx": 11713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911330.441, "dur": 9.048,
    "args": {
      "Python parent id": 8810, "Python id": 8814, "finished": true, "Ev Idx": 11714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911330.560, "dur": 0.067,
    "args": {
      "Python parent id": 8814, "Python id": 8815, "finished": true, "Ev Idx": 11715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911330.763, "dur": 0.063,
    "args": {
      "Python parent id": 8814, "Python id": 8816, "finished": true, "Ev Idx": 11716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911330.977, "dur": 0.059,
    "args": {
      "Python parent id": 8814, "Python id": 8817, "finished": true, "Ev Idx": 11717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911331.476, "dur": 0.040,
    "args": {
      "Python parent id": 8814, "Python id": 8818, "finished": true, "Ev Idx": 11718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911331.765, "dur": 0.080,
    "args": {
      "Python parent id": 8814, "Python id": 8819, "finished": true, "Ev Idx": 11719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911332.225, "dur": 7.190,
    "args": {
      "Python parent id": 8814, "Python id": 8820, "finished": true, "Ev Idx": 11720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911332.312, "dur": 0.063,
    "args": {
      "Python parent id": 8820, "Python id": 8821, "finished": true, "Ev Idx": 11721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911332.527, "dur": 0.062,
    "args": {
      "Python parent id": 8820, "Python id": 8822, "finished": true, "Ev Idx": 11722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911332.751, "dur": 0.070,
    "args": {
      "Python parent id": 8820, "Python id": 8823, "finished": true, "Ev Idx": 11723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911333.156, "dur": 0.036,
    "args": {
      "Python parent id": 8820, "Python id": 8824, "finished": true, "Ev Idx": 11724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911333.519, "dur": 5.784,
    "args": {
      "Python parent id": 8820, "Python id": 8825, "finished": true, "Ev Idx": 11725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911333.793, "dur": 5.437,
    "args": {
      "Python parent id": 8825, "Python id": 8826, "finished": true, "Ev Idx": 11726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911333.935, "dur": 3.486,
    "args": {
      "Python parent id": 8826, "Python id": 8827, "finished": true, "Ev Idx": 11727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911334.138, "dur": 3.242,
    "args": {
      "Python parent id": 8827, "Python id": 8828, "finished": true, "Ev Idx": 11728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911334.257, "dur": 0.289,
    "args": {
      "Python parent id": 8828, "Python id": 8829, "finished": true, "Ev Idx": 11729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911334.395, "dur": 0.085,
    "args": {
      "Python parent id": 8829, "Python id": 8830, "finished": true, "Ev Idx": 11730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911334.738, "dur": 2.337,
    "args": {
      "Python parent id": 8828, "Python id": 8831, "finished": true, "Ev Idx": 11731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911334.899, "dur": 2.092,
    "args": {
      "Python parent id": 8831, "Python id": 8832, "finished": true, "Ev Idx": 11732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911335.148, "dur": 1.767,
    "args": {
      "Python parent id": 8832, "Python id": 8833, "finished": true, "Ev Idx": 11733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911335.286, "dur": 1.358,
    "args": {
      "Python parent id": 8833, "Python id": 8834, "finished": true, "Ev Idx": 11734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911335.519, "dur": 0.582,
    "args": {
      "Python parent id": 8834, "Python id": 8835, "finished": true, "Ev Idx": 11735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911335.598, "dur": 0.057,
    "args": {
      "Python parent id": 8835, "Python id": 8836, "finished": true, "Ev Idx": 11736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911335.863, "dur": 0.176,
    "args": {
      "Python parent id": 8835, "Python id": 8837, "finished": true, "Ev Idx": 11737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911337.254, "dur": 0.072,
    "args": {
      "Python parent id": 8828, "Python id": 8838, "finished": true, "Ev Idx": 11738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911337.610, "dur": 0.074,
    "args": {
      "Python parent id": 8826, "Python id": 8839, "finished": true, "Ev Idx": 11739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911337.950, "dur": 1.239,
    "args": {
      "Python parent id": 8826, "Python id": 8840, "finished": true, "Ev Idx": 11740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911338.114, "dur": 1.029,
    "args": {
      "Python parent id": 8840, "Python id": 8841, "finished": true, "Ev Idx": 11741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911338.259, "dur": 0.407,
    "args": {
      "Python parent id": 8841, "Python id": 8842, "finished": true, "Ev Idx": 11742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911338.421, "dur": 0.184,
    "args": {
      "Python parent id": 8842, "Python id": 8843, "finished": true, "Ev Idx": 11743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911338.513, "dur": 0.060,
    "args": {
      "Python parent id": 8843, "Python id": 8844, "finished": true, "Ev Idx": 11744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911338.819, "dur": 0.296,
    "args": {
      "Python parent id": 8841, "Python id": 8845, "finished": true, "Ev Idx": 11745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911339.627, "dur": 0.158,
    "args": {
      "Python parent id": 8810, "Python id": 8846, "finished": true, "Ev Idx": 11746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911340.463, "dur": 0.955,
    "args": {
      "Python parent id": 8810, "Python id": 8847, "finished": true, "Ev Idx": 11747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911340.882, "dur": 0.473,
    "args": {
      "Python parent id": 8847, "Python id": 8848, "finished": true, "Ev Idx": 11748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911341.879, "dur": 4.300,
    "args": {
      "Python parent id": 8809, "Python id": 8849, "finished": true, "Ev Idx": 11749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911347.320, "dur": 95.937,
    "args": {
      "Python parent id": 12, "Python id": 8850, "finished": true, "Ev Idx": 11750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911347.782, "dur": 0.523,
    "args": {
      "Python parent id": 8850, "Python id": 8851, "finished": true, "Ev Idx": 11751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911348.118, "dur": 0.123,
    "args": {
      "Python parent id": 8851, "Python id": 8852, "finished": true, "Ev Idx": 11752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911348.639, "dur": 0.343,
    "args": {
      "Python parent id": 8850, "Python id": 8853, "finished": true, "Ev Idx": 11753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911349.236, "dur": 0.287,
    "args": {
      "Python parent id": 8850, "Python id": 8854, "finished": true, "Ev Idx": 11754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911349.778, "dur": 1.760,
    "args": {
      "Python parent id": 8850, "Python id": 8855, "finished": true, "Ev Idx": 11755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911352.300, "dur": 90.807,
    "args": {
      "Python parent id": 8850, "Python id": 8856, "finished": true, "Ev Idx": 11756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911353.394, "dur": 89.542,
    "args": {
      "Python parent id": 8856, "Python id": 8857, "finished": true, "Ev Idx": 11757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911353.792, "dur": 89.063,
    "args": {
      "Python parent id": 8857, "Python id": 8858, "finished": true, "Ev Idx": 11758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911359.946, "dur": 81.323,
    "args": {
      "Python parent id": 8858, "Python id": 8859, "finished": true, "Ev Idx": 11759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911360.347, "dur": 80.789,
    "args": {
      "Python parent id": 8859, "Python id": 8860, "finished": true, "Ev Idx": 11760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911361.025, "dur": 80.038,
    "args": {
      "Python parent id": 8860, "Python id": 8861, "finished": true, "Ev Idx": 11761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911361.628, "dur": 0.492,
    "args": {
      "Python parent id": 8861, "Python id": 8862, "finished": true, "Ev Idx": 11762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911361.981, "dur": 0.085,
    "args": {
      "Python parent id": 8862, "Python id": 8863, "finished": true, "Ev Idx": 11763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911362.529, "dur": 0.377,
    "args": {
      "Python parent id": 8861, "Python id": 8864, "finished": true, "Ev Idx": 11764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911362.816, "dur": 0.042,
    "args": {
      "Python parent id": 8864, "Python id": 8865, "finished": true, "Ev Idx": 11765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911363.282, "dur": 77.677,
    "args": {
      "Python parent id": 8861, "Python id": 8866, "finished": true, "Ev Idx": 11766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911364.143, "dur": 75.428,
    "args": {
      "Python parent id": 8866, "Python id": 8867, "finished": true, "Ev Idx": 11767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911371.784, "dur": 18.571,
    "args": {
      "Python parent id": 8867, "Python id": 8868, "finished": true, "Ev Idx": 11768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911371.926, "dur": 0.137,
    "args": {
      "Python parent id": 8868, "Python id": 8869, "finished": true, "Ev Idx": 11769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911372.421, "dur": 17.791,
    "args": {
      "Python parent id": 8868, "Python id": 8870, "finished": true, "Ev Idx": 11770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911372.549, "dur": 7.384,
    "args": {
      "Python parent id": 8870, "Python id": 8871, "finished": true, "Ev Idx": 11771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911399.033, "dur": 7.447,
    "args": {
      "Python parent id": 8867, "Python id": 8872, "finished": true, "Ev Idx": 11772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911414.734, "dur": 8.843,
    "args": {
      "Python parent id": 8867, "Python id": 8873, "finished": true, "Ev Idx": 11773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911444.466, "dur": 17.671,
    "args": {
      "Python parent id": 12, "Python id": 8874, "finished": true, "Ev Idx": 11774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911444.828, "dur": 13.643,
    "args": {
      "Python parent id": 8874, "Python id": 8875, "finished": true, "Ev Idx": 11775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911445.045, "dur": 0.545,
    "args": {
      "Python parent id": 8875, "Python id": 8876, "finished": true, "Ev Idx": 11776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911445.257, "dur": 0.264,
    "args": {
      "Python parent id": 8876, "Python id": 8877, "finished": true, "Ev Idx": 11777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911445.376, "dur": 0.110,
    "args": {
      "Python parent id": 8877, "Python id": 8878, "finished": true, "Ev Idx": 11778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911446.059, "dur": 9.939,
    "args": {
      "Python parent id": 8875, "Python id": 8879, "finished": true, "Ev Idx": 11779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911446.161, "dur": 0.095,
    "args": {
      "Python parent id": 8879, "Python id": 8880, "finished": true, "Ev Idx": 11780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911446.376, "dur": 0.056,
    "args": {
      "Python parent id": 8879, "Python id": 8881, "finished": true, "Ev Idx": 11781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911446.583, "dur": 0.060,
    "args": {
      "Python parent id": 8879, "Python id": 8882, "finished": true, "Ev Idx": 11782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911447.083, "dur": 0.037,
    "args": {
      "Python parent id": 8879, "Python id": 8883, "finished": true, "Ev Idx": 11783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911447.411, "dur": 0.084,
    "args": {
      "Python parent id": 8879, "Python id": 8884, "finished": true, "Ev Idx": 11784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911447.862, "dur": 8.059,
    "args": {
      "Python parent id": 8879, "Python id": 8885, "finished": true, "Ev Idx": 11785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911448.005, "dur": 0.060,
    "args": {
      "Python parent id": 8885, "Python id": 8886, "finished": true, "Ev Idx": 11786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911448.228, "dur": 0.052,
    "args": {
      "Python parent id": 8885, "Python id": 8887, "finished": true, "Ev Idx": 11787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911448.484, "dur": 0.064,
    "args": {
      "Python parent id": 8885, "Python id": 8888, "finished": true, "Ev Idx": 11788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911448.957, "dur": 0.049,
    "args": {
      "Python parent id": 8885, "Python id": 8889, "finished": true, "Ev Idx": 11789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911449.360, "dur": 6.484,
    "args": {
      "Python parent id": 8885, "Python id": 8890, "finished": true, "Ev Idx": 11790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911449.664, "dur": 6.105,
    "args": {
      "Python parent id": 8890, "Python id": 8891, "finished": true, "Ev Idx": 11791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911449.858, "dur": 3.976,
    "args": {
      "Python parent id": 8891, "Python id": 8892, "finished": true, "Ev Idx": 11792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911450.090, "dur": 3.674,
    "args": {
      "Python parent id": 8892, "Python id": 8893, "finished": true, "Ev Idx": 11793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911450.251, "dur": 0.320,
    "args": {
      "Python parent id": 8893, "Python id": 8894, "finished": true, "Ev Idx": 11794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911450.412, "dur": 0.063,
    "args": {
      "Python parent id": 8894, "Python id": 8895, "finished": true, "Ev Idx": 11795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911450.772, "dur": 2.568,
    "args": {
      "Python parent id": 8893, "Python id": 8896, "finished": true, "Ev Idx": 11796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911451.024, "dur": 2.197,
    "args": {
      "Python parent id": 8896, "Python id": 8897, "finished": true, "Ev Idx": 11797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911451.247, "dur": 1.890,
    "args": {
      "Python parent id": 8897, "Python id": 8898, "finished": true, "Ev Idx": 11798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911451.401, "dur": 1.480,
    "args": {
      "Python parent id": 8898, "Python id": 8899, "finished": true, "Ev Idx": 11799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911451.664, "dur": 0.579,
    "args": {
      "Python parent id": 8899, "Python id": 8900, "finished": true, "Ev Idx": 11800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911451.769, "dur": 0.063,
    "args": {
      "Python parent id": 8900, "Python id": 8901, "finished": true, "Ev Idx": 11801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911452.029, "dur": 0.167,
    "args": {
      "Python parent id": 8900, "Python id": 8902, "finished": true, "Ev Idx": 11802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911453.543, "dur": 0.159,
    "args": {
      "Python parent id": 8893, "Python id": 8903, "finished": true, "Ev Idx": 11803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.018, "dur": 0.074,
    "args": {
      "Python parent id": 8891, "Python id": 8904, "finished": true, "Ev Idx": 11804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.380, "dur": 1.311,
    "args": {
      "Python parent id": 8891, "Python id": 8905, "finished": true, "Ev Idx": 11805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.611, "dur": 1.040,
    "args": {
      "Python parent id": 8905, "Python id": 8906, "finished": true, "Ev Idx": 11806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.744, "dur": 0.407,
    "args": {
      "Python parent id": 8906, "Python id": 8907, "finished": true, "Ev Idx": 11807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.883, "dur": 0.209,
    "args": {
      "Python parent id": 8907, "Python id": 8908, "finished": true, "Ev Idx": 11808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911454.983, "dur": 0.079,
    "args": {
      "Python parent id": 8908, "Python id": 8909, "finished": true, "Ev Idx": 11809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911455.294, "dur": 0.310,
    "args": {
      "Python parent id": 8906, "Python id": 8910, "finished": true, "Ev Idx": 11810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911456.137, "dur": 0.244,
    "args": {
      "Python parent id": 8875, "Python id": 8911, "finished": true, "Ev Idx": 11811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911457.122, "dur": 1.205,
    "args": {
      "Python parent id": 8875, "Python id": 8912, "finished": true, "Ev Idx": 11812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911457.634, "dur": 0.639,
    "args": {
      "Python parent id": 8912, "Python id": 8913, "finished": true, "Ev Idx": 11813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911458.797, "dur": 3.225,
    "args": {
      "Python parent id": 8874, "Python id": 8914, "finished": true, "Ev Idx": 11814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911462.804, "dur": 7.829,
    "args": {
      "Python parent id": 12, "Python id": 8915, "finished": true, "Ev Idx": 11815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911471.126, "dur": 0.114,
    "args": {
      "Python parent id": 12, "Python id": 8916, "finished": true, "Ev Idx": 11816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911471.897, "dur": 0.246,
    "args": {
      "Python parent id": 12, "Python id": 8917, "finished": true, "Ev Idx": 11817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911472.693, "dur": 0.093,
    "args": {
      "Python parent id": 12, "Python id": 8918, "finished": true, "Ev Idx": 11818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911473.225, "dur": 10.784,
    "args": {
      "Python parent id": 12, "Python id": 8919, "finished": true, "Ev Idx": 11819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911485.444, "dur": 16.159,
    "args": {
      "Python parent id": 12, "Python id": 8920, "finished": true, "Ev Idx": 11820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911485.790, "dur": 11.809,
    "args": {
      "Python parent id": 8920, "Python id": 8921, "finished": true, "Ev Idx": 11821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911485.951, "dur": 0.509,
    "args": {
      "Python parent id": 8921, "Python id": 8922, "finished": true, "Ev Idx": 11822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911486.120, "dur": 0.236,
    "args": {
      "Python parent id": 8922, "Python id": 8923, "finished": true, "Ev Idx": 11823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911486.223, "dur": 0.098,
    "args": {
      "Python parent id": 8923, "Python id": 8924, "finished": true, "Ev Idx": 11824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911486.706, "dur": 8.867,
    "args": {
      "Python parent id": 8921, "Python id": 8925, "finished": true, "Ev Idx": 11825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911486.834, "dur": 0.111,
    "args": {
      "Python parent id": 8925, "Python id": 8926, "finished": true, "Ev Idx": 11826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911487.043, "dur": 0.064,
    "args": {
      "Python parent id": 8925, "Python id": 8927, "finished": true, "Ev Idx": 11827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911487.287, "dur": 0.062,
    "args": {
      "Python parent id": 8925, "Python id": 8928, "finished": true, "Ev Idx": 11828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911487.801, "dur": 0.043,
    "args": {
      "Python parent id": 8925, "Python id": 8929, "finished": true, "Ev Idx": 11829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911488.151, "dur": 0.086,
    "args": {
      "Python parent id": 8925, "Python id": 8930, "finished": true, "Ev Idx": 11830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911488.591, "dur": 6.914,
    "args": {
      "Python parent id": 8925, "Python id": 8931, "finished": true, "Ev Idx": 11831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911488.691, "dur": 0.068,
    "args": {
      "Python parent id": 8931, "Python id": 8932, "finished": true, "Ev Idx": 11832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911488.912, "dur": 0.062,
    "args": {
      "Python parent id": 8931, "Python id": 8933, "finished": true, "Ev Idx": 11833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911489.171, "dur": 0.069,
    "args": {
      "Python parent id": 8931, "Python id": 8934, "finished": true, "Ev Idx": 11834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911489.624, "dur": 0.037,
    "args": {
      "Python parent id": 8931, "Python id": 8935, "finished": true, "Ev Idx": 11835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.000, "dur": 5.424,
    "args": {
      "Python parent id": 8931, "Python id": 8936, "finished": true, "Ev Idx": 11836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.258, "dur": 5.106,
    "args": {
      "Python parent id": 8936, "Python id": 8937, "finished": true, "Ev Idx": 11837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.399, "dur": 3.234,
    "args": {
      "Python parent id": 8937, "Python id": 8938, "finished": true, "Ev Idx": 11838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.628, "dur": 2.958,
    "args": {
      "Python parent id": 8938, "Python id": 8939, "finished": true, "Ev Idx": 11839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.781, "dur": 0.271,
    "args": {
      "Python parent id": 8939, "Python id": 8940, "finished": true, "Ev Idx": 11840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911490.905, "dur": 0.067,
    "args": {
      "Python parent id": 8940, "Python id": 8941, "finished": true, "Ev Idx": 11841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911491.200, "dur": 2.029,
    "args": {
      "Python parent id": 8939, "Python id": 8942, "finished": true, "Ev Idx": 11842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911491.409, "dur": 1.758,
    "args": {
      "Python parent id": 8942, "Python id": 8943, "finished": true, "Ev Idx": 11843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911491.607, "dur": 1.512,
    "args": {
      "Python parent id": 8943, "Python id": 8944, "finished": true, "Ev Idx": 11844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911491.731, "dur": 1.170,
    "args": {
      "Python parent id": 8944, "Python id": 8945, "finished": true, "Ev Idx": 11845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911491.923, "dur": 0.515,
    "args": {
      "Python parent id": 8945, "Python id": 8946, "finished": true, "Ev Idx": 11846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911492.040, "dur": 0.048,
    "args": {
      "Python parent id": 8946, "Python id": 8947, "finished": true, "Ev Idx": 11847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911492.267, "dur": 0.126,
    "args": {
      "Python parent id": 8946, "Python id": 8948, "finished": true, "Ev Idx": 11848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911493.416, "dur": 0.113,
    "args": {
      "Python parent id": 8939, "Python id": 8949, "finished": true, "Ev Idx": 11849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911493.845, "dur": 0.078,
    "args": {
      "Python parent id": 8937, "Python id": 8950, "finished": true, "Ev Idx": 11850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.187, "dur": 1.108,
    "args": {
      "Python parent id": 8937, "Python id": 8951, "finished": true, "Ev Idx": 11851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.346, "dur": 0.903,
    "args": {
      "Python parent id": 8951, "Python id": 8952, "finished": true, "Ev Idx": 11852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.481, "dur": 0.370,
    "args": {
      "Python parent id": 8952, "Python id": 8953, "finished": true, "Ev Idx": 11853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.612, "dur": 0.181,
    "args": {
      "Python parent id": 8953, "Python id": 8954, "finished": true, "Ev Idx": 11854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.704, "dur": 0.056,
    "args": {
      "Python parent id": 8954, "Python id": 8955, "finished": true, "Ev Idx": 11855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911494.995, "dur": 0.220,
    "args": {
      "Python parent id": 8952, "Python id": 8956, "finished": true, "Ev Idx": 11856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911495.737, "dur": 0.162,
    "args": {
      "Python parent id": 8921, "Python id": 8957, "finished": true, "Ev Idx": 11857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911496.507, "dur": 0.957,
    "args": {
      "Python parent id": 8921, "Python id": 8958, "finished": true, "Ev Idx": 11858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911496.974, "dur": 0.441,
    "args": {
      "Python parent id": 8958, "Python id": 8959, "finished": true, "Ev Idx": 11859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911497.911, "dur": 3.642,
    "args": {
      "Python parent id": 8920, "Python id": 8960, "finished": true, "Ev Idx": 11860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911502.575, "dur": 92.168,
    "args": {
      "Python parent id": 12, "Python id": 8961, "finished": true, "Ev Idx": 11861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911502.946, "dur": 0.443,
    "args": {
      "Python parent id": 8961, "Python id": 8962, "finished": true, "Ev Idx": 11862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911503.233, "dur": 0.108,
    "args": {
      "Python parent id": 8962, "Python id": 8963, "finished": true, "Ev Idx": 11863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911503.724, "dur": 0.429,
    "args": {
      "Python parent id": 8961, "Python id": 8964, "finished": true, "Ev Idx": 11864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911504.404, "dur": 0.311,
    "args": {
      "Python parent id": 8961, "Python id": 8965, "finished": true, "Ev Idx": 11865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911504.974, "dur": 1.632,
    "args": {
      "Python parent id": 8961, "Python id": 8966, "finished": true, "Ev Idx": 11866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911507.296, "dur": 87.351,
    "args": {
      "Python parent id": 8961, "Python id": 8967, "finished": true, "Ev Idx": 11867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911508.280, "dur": 86.151,
    "args": {
      "Python parent id": 8967, "Python id": 8968, "finished": true, "Ev Idx": 11868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911508.640, "dur": 85.713,
    "args": {
      "Python parent id": 8968, "Python id": 8969, "finished": true, "Ev Idx": 11869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911514.264, "dur": 78.708,
    "args": {
      "Python parent id": 8969, "Python id": 8970, "finished": true, "Ev Idx": 11870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911514.699, "dur": 78.090,
    "args": {
      "Python parent id": 8970, "Python id": 8971, "finished": true, "Ev Idx": 11871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911515.368, "dur": 77.347,
    "args": {
      "Python parent id": 8971, "Python id": 8972, "finished": true, "Ev Idx": 11872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911515.952, "dur": 0.403,
    "args": {
      "Python parent id": 8972, "Python id": 8973, "finished": true, "Ev Idx": 11873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911516.210, "dur": 0.094,
    "args": {
      "Python parent id": 8973, "Python id": 8974, "finished": true, "Ev Idx": 11874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911516.666, "dur": 0.319,
    "args": {
      "Python parent id": 8972, "Python id": 8975, "finished": true, "Ev Idx": 11875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911516.894, "dur": 0.039,
    "args": {
      "Python parent id": 8975, "Python id": 8976, "finished": true, "Ev Idx": 11876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911517.327, "dur": 75.277,
    "args": {
      "Python parent id": 8972, "Python id": 8977, "finished": true, "Ev Idx": 11877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911518.013, "dur": 73.244,
    "args": {
      "Python parent id": 8977, "Python id": 8978, "finished": true, "Ev Idx": 11878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911525.341, "dur": 18.101,
    "args": {
      "Python parent id": 8978, "Python id": 8979, "finished": true, "Ev Idx": 11879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911525.514, "dur": 0.125,
    "args": {
      "Python parent id": 8979, "Python id": 8980, "finished": true, "Ev Idx": 11880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911525.937, "dur": 17.388,
    "args": {
      "Python parent id": 8979, "Python id": 8981, "finished": true, "Ev Idx": 11881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911526.068, "dur": 7.693,
    "args": {
      "Python parent id": 8981, "Python id": 8982, "finished": true, "Ev Idx": 11882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911550.756, "dur": 7.299,
    "args": {
      "Python parent id": 8978, "Python id": 8983, "finished": true, "Ev Idx": 11883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911567.365, "dur": 8.760,
    "args": {
      "Python parent id": 8978, "Python id": 8984, "finished": true, "Ev Idx": 11884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911595.953, "dur": 22.843,
    "args": {
      "Python parent id": 12, "Python id": 8985, "finished": true, "Ev Idx": 11885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911596.294, "dur": 13.452,
    "args": {
      "Python parent id": 8985, "Python id": 8986, "finished": true, "Ev Idx": 11886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911596.490, "dur": 0.533,
    "args": {
      "Python parent id": 8986, "Python id": 8987, "finished": true, "Ev Idx": 11887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911596.685, "dur": 0.259,
    "args": {
      "Python parent id": 8987, "Python id": 8988, "finished": true, "Ev Idx": 11888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911596.843, "dur": 0.066,
    "args": {
      "Python parent id": 8988, "Python id": 8989, "finished": true, "Ev Idx": 11889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911597.401, "dur": 9.896,
    "args": {
      "Python parent id": 8986, "Python id": 8990, "finished": true, "Ev Idx": 11890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911597.559, "dur": 0.066,
    "args": {
      "Python parent id": 8990, "Python id": 8991, "finished": true, "Ev Idx": 11891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911597.765, "dur": 0.058,
    "args": {
      "Python parent id": 8990, "Python id": 8992, "finished": true, "Ev Idx": 11892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911597.978, "dur": 0.068,
    "args": {
      "Python parent id": 8990, "Python id": 8993, "finished": true, "Ev Idx": 11893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911598.543, "dur": 0.056,
    "args": {
      "Python parent id": 8990, "Python id": 8994, "finished": true, "Ev Idx": 11894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911598.926, "dur": 0.090,
    "args": {
      "Python parent id": 8990, "Python id": 8995, "finished": true, "Ev Idx": 11895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911599.334, "dur": 7.889,
    "args": {
      "Python parent id": 8990, "Python id": 8996, "finished": true, "Ev Idx": 11896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911599.435, "dur": 0.062,
    "args": {
      "Python parent id": 8996, "Python id": 8997, "finished": true, "Ev Idx": 11897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911599.683, "dur": 0.064,
    "args": {
      "Python parent id": 8996, "Python id": 8998, "finished": true, "Ev Idx": 11898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911599.921, "dur": 0.067,
    "args": {
      "Python parent id": 8996, "Python id": 8999, "finished": true, "Ev Idx": 11899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911600.326, "dur": 0.040,
    "args": {
      "Python parent id": 8996, "Python id": 9000, "finished": true, "Ev Idx": 11900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911600.725, "dur": 6.374,
    "args": {
      "Python parent id": 8996, "Python id": 9001, "finished": true, "Ev Idx": 11901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911600.951, "dur": 6.086,
    "args": {
      "Python parent id": 9001, "Python id": 9002, "finished": true, "Ev Idx": 11902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911601.084, "dur": 3.808,
    "args": {
      "Python parent id": 9002, "Python id": 9003, "finished": true, "Ev Idx": 11903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911601.298, "dur": 3.545,
    "args": {
      "Python parent id": 9003, "Python id": 9004, "finished": true, "Ev Idx": 11904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911601.438, "dur": 0.297,
    "args": {
      "Python parent id": 9004, "Python id": 9005, "finished": true, "Ev Idx": 11905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911601.595, "dur": 0.064,
    "args": {
      "Python parent id": 9005, "Python id": 9006, "finished": true, "Ev Idx": 11906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911601.935, "dur": 2.467,
    "args": {
      "Python parent id": 9004, "Python id": 9007, "finished": true, "Ev Idx": 11907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911602.207, "dur": 2.116,
    "args": {
      "Python parent id": 9007, "Python id": 9008, "finished": true, "Ev Idx": 11908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911602.433, "dur": 1.824,
    "args": {
      "Python parent id": 9008, "Python id": 9009, "finished": true, "Ev Idx": 11909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911602.583, "dur": 1.407,
    "args": {
      "Python parent id": 9009, "Python id": 9010, "finished": true, "Ev Idx": 11910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911602.797, "dur": 0.586,
    "args": {
      "Python parent id": 9010, "Python id": 9011, "finished": true, "Ev Idx": 11911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911602.905, "dur": 0.046,
    "args": {
      "Python parent id": 9011, "Python id": 9012, "finished": true, "Ev Idx": 11912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911603.162, "dur": 0.141,
    "args": {
      "Python parent id": 9011, "Python id": 9013, "finished": true, "Ev Idx": 11913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911604.602, "dur": 0.174,
    "args": {
      "Python parent id": 9004, "Python id": 9014, "finished": true, "Ev Idx": 11914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911605.048, "dur": 0.085,
    "args": {
      "Python parent id": 9002, "Python id": 9015, "finished": true, "Ev Idx": 11915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911605.446, "dur": 1.543,
    "args": {
      "Python parent id": 9002, "Python id": 9016, "finished": true, "Ev Idx": 11916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911605.665, "dur": 1.281,
    "args": {
      "Python parent id": 9016, "Python id": 9017, "finished": true, "Ev Idx": 11917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911605.851, "dur": 0.519,
    "args": {
      "Python parent id": 9017, "Python id": 9018, "finished": true, "Ev Idx": 11918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911606.063, "dur": 0.242,
    "args": {
      "Python parent id": 9018, "Python id": 9019, "finished": true, "Ev Idx": 11919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911606.202, "dur": 0.067,
    "args": {
      "Python parent id": 9019, "Python id": 9020, "finished": true, "Ev Idx": 11920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911606.553, "dur": 0.341,
    "args": {
      "Python parent id": 9017, "Python id": 9021, "finished": true, "Ev Idx": 11921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911607.430, "dur": 0.269,
    "args": {
      "Python parent id": 8986, "Python id": 9022, "finished": true, "Ev Idx": 11922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911608.394, "dur": 1.185,
    "args": {
      "Python parent id": 8986, "Python id": 9023, "finished": true, "Ev Idx": 11923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911608.928, "dur": 0.553,
    "args": {
      "Python parent id": 9023, "Python id": 9024, "finished": true, "Ev Idx": 11924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911610.170, "dur": 8.481,
    "args": {
      "Python parent id": 8985, "Python id": 9025, "finished": true, "Ev Idx": 11925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911619.374, "dur": 8.503,
    "args": {
      "Python parent id": 12, "Python id": 9026, "finished": true, "Ev Idx": 11926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911628.405, "dur": 0.097,
    "args": {
      "Python parent id": 12, "Python id": 9027, "finished": true, "Ev Idx": 11927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911629.103, "dur": 0.241,
    "args": {
      "Python parent id": 12, "Python id": 9028, "finished": true, "Ev Idx": 11928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911629.858, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 9029, "finished": true, "Ev Idx": 11929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911630.347, "dur": 10.704,
    "args": {
      "Python parent id": 12, "Python id": 9030, "finished": true, "Ev Idx": 11930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911642.592, "dur": 16.029,
    "args": {
      "Python parent id": 12, "Python id": 9031, "finished": true, "Ev Idx": 11931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911642.908, "dur": 11.761,
    "args": {
      "Python parent id": 9031, "Python id": 9032, "finished": true, "Ev Idx": 11932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911643.072, "dur": 0.524,
    "args": {
      "Python parent id": 9032, "Python id": 9033, "finished": true, "Ev Idx": 11933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911643.240, "dur": 0.255,
    "args": {
      "Python parent id": 9033, "Python id": 9034, "finished": true, "Ev Idx": 11934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911643.359, "dur": 0.099,
    "args": {
      "Python parent id": 9034, "Python id": 9035, "finished": true, "Ev Idx": 11935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911643.925, "dur": 8.822,
    "args": {
      "Python parent id": 9032, "Python id": 9036, "finished": true, "Ev Idx": 11936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911644.007, "dur": 0.070,
    "args": {
      "Python parent id": 9036, "Python id": 9037, "finished": true, "Ev Idx": 11937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911644.193, "dur": 0.061,
    "args": {
      "Python parent id": 9036, "Python id": 9038, "finished": true, "Ev Idx": 11938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911644.454, "dur": 0.061,
    "args": {
      "Python parent id": 9036, "Python id": 9039, "finished": true, "Ev Idx": 11939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911644.965, "dur": 0.065,
    "args": {
      "Python parent id": 9036, "Python id": 9040, "finished": true, "Ev Idx": 11940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911645.326, "dur": 0.084,
    "args": {
      "Python parent id": 9036, "Python id": 9041, "finished": true, "Ev Idx": 11941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911645.777, "dur": 6.912,
    "args": {
      "Python parent id": 9036, "Python id": 9042, "finished": true, "Ev Idx": 11942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911645.917, "dur": 0.066,
    "args": {
      "Python parent id": 9042, "Python id": 9043, "finished": true, "Ev Idx": 11943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911646.135, "dur": 0.068,
    "args": {
      "Python parent id": 9042, "Python id": 9044, "finished": true, "Ev Idx": 11944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911646.376, "dur": 0.086,
    "args": {
      "Python parent id": 9042, "Python id": 9045, "finished": true, "Ev Idx": 11945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911646.823, "dur": 0.028,
    "args": {
      "Python parent id": 9042, "Python id": 9046, "finished": true, "Ev Idx": 11946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.173, "dur": 5.410,
    "args": {
      "Python parent id": 9042, "Python id": 9047, "finished": true, "Ev Idx": 11947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.362, "dur": 5.165,
    "args": {
      "Python parent id": 9047, "Python id": 9048, "finished": true, "Ev Idx": 11948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.497, "dur": 3.306,
    "args": {
      "Python parent id": 9048, "Python id": 9049, "finished": true, "Ev Idx": 11949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.680, "dur": 3.082,
    "args": {
      "Python parent id": 9049, "Python id": 9050, "finished": true, "Ev Idx": 11950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.842, "dur": 0.246,
    "args": {
      "Python parent id": 9050, "Python id": 9051, "finished": true, "Ev Idx": 11951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911647.968, "dur": 0.066,
    "args": {
      "Python parent id": 9051, "Python id": 9052, "finished": true, "Ev Idx": 11952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911648.226, "dur": 2.160,
    "args": {
      "Python parent id": 9050, "Python id": 9053, "finished": true, "Ev Idx": 11953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911648.431, "dur": 1.887,
    "args": {
      "Python parent id": 9053, "Python id": 9054, "finished": true, "Ev Idx": 11954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911648.638, "dur": 1.620,
    "args": {
      "Python parent id": 9054, "Python id": 9055, "finished": true, "Ev Idx": 11955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911648.814, "dur": 1.176,
    "args": {
      "Python parent id": 9055, "Python id": 9056, "finished": true, "Ev Idx": 11956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911648.996, "dur": 0.481,
    "args": {
      "Python parent id": 9056, "Python id": 9057, "finished": true, "Ev Idx": 11957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911649.076, "dur": 0.051,
    "args": {
      "Python parent id": 9057, "Python id": 9058, "finished": true, "Ev Idx": 11958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911649.305, "dur": 0.115,
    "args": {
      "Python parent id": 9057, "Python id": 9059, "finished": true, "Ev Idx": 11959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911650.600, "dur": 0.109,
    "args": {
      "Python parent id": 9050, "Python id": 9060, "finished": true, "Ev Idx": 11960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911650.950, "dur": 0.080,
    "args": {
      "Python parent id": 9048, "Python id": 9061, "finished": true, "Ev Idx": 11961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911651.282, "dur": 1.201,
    "args": {
      "Python parent id": 9048, "Python id": 9062, "finished": true, "Ev Idx": 11962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911651.423, "dur": 1.013,
    "args": {
      "Python parent id": 9062, "Python id": 9063, "finished": true, "Ev Idx": 11963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911651.569, "dur": 0.388,
    "args": {
      "Python parent id": 9063, "Python id": 9064, "finished": true, "Ev Idx": 11964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911651.709, "dur": 0.191,
    "args": {
      "Python parent id": 9064, "Python id": 9065, "finished": true, "Ev Idx": 11965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911651.812, "dur": 0.056,
    "args": {
      "Python parent id": 9065, "Python id": 9066, "finished": true, "Ev Idx": 11966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911652.093, "dur": 0.314,
    "args": {
      "Python parent id": 9063, "Python id": 9067, "finished": true, "Ev Idx": 11967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911652.898, "dur": 0.132,
    "args": {
      "Python parent id": 9032, "Python id": 9068, "finished": true, "Ev Idx": 11968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911653.601, "dur": 0.937,
    "args": {
      "Python parent id": 9032, "Python id": 9069, "finished": true, "Ev Idx": 11969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911654.015, "dur": 0.471,
    "args": {
      "Python parent id": 9069, "Python id": 9070, "finished": true, "Ev Idx": 11970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911654.961, "dur": 3.583,
    "args": {
      "Python parent id": 9031, "Python id": 9071, "finished": true, "Ev Idx": 11971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911659.504, "dur": 96.706,
    "args": {
      "Python parent id": 12, "Python id": 9072, "finished": true, "Ev Idx": 11972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911659.859, "dur": 0.500,
    "args": {
      "Python parent id": 9072, "Python id": 9073, "finished": true, "Ev Idx": 11973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911660.158, "dur": 0.125,
    "args": {
      "Python parent id": 9073, "Python id": 9074, "finished": true, "Ev Idx": 11974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911660.652, "dur": 0.390,
    "args": {
      "Python parent id": 9072, "Python id": 9075, "finished": true, "Ev Idx": 11975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911661.287, "dur": 0.266,
    "args": {
      "Python parent id": 9072, "Python id": 9076, "finished": true, "Ev Idx": 11976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911661.822, "dur": 1.540,
    "args": {
      "Python parent id": 9072, "Python id": 9077, "finished": true, "Ev Idx": 11977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911663.954, "dur": 92.130,
    "args": {
      "Python parent id": 9072, "Python id": 9078, "finished": true, "Ev Idx": 11978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911664.858, "dur": 91.035,
    "args": {
      "Python parent id": 9078, "Python id": 9079, "finished": true, "Ev Idx": 11979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911665.106, "dur": 90.686,
    "args": {
      "Python parent id": 9079, "Python id": 9080, "finished": true, "Ev Idx": 11980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911670.529, "dur": 83.895,
    "args": {
      "Python parent id": 9080, "Python id": 9081, "finished": true, "Ev Idx": 11981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911670.858, "dur": 83.486,
    "args": {
      "Python parent id": 9081, "Python id": 9082, "finished": true, "Ev Idx": 11982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911671.513, "dur": 82.683,
    "args": {
      "Python parent id": 9082, "Python id": 9083, "finished": true, "Ev Idx": 11983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911672.150, "dur": 0.447,
    "args": {
      "Python parent id": 9083, "Python id": 9084, "finished": true, "Ev Idx": 11984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911672.416, "dur": 0.082,
    "args": {
      "Python parent id": 9084, "Python id": 9085, "finished": true, "Ev Idx": 11985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911672.919, "dur": 0.327,
    "args": {
      "Python parent id": 9083, "Python id": 9086, "finished": true, "Ev Idx": 11986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911673.143, "dur": 0.048,
    "args": {
      "Python parent id": 9086, "Python id": 9087, "finished": true, "Ev Idx": 11987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911673.614, "dur": 80.449,
    "args": {
      "Python parent id": 9083, "Python id": 9088, "finished": true, "Ev Idx": 11988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911674.284, "dur": 78.438,
    "args": {
      "Python parent id": 9088, "Python id": 9089, "finished": true, "Ev Idx": 11989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911681.587, "dur": 20.003,
    "args": {
      "Python parent id": 9089, "Python id": 9090, "finished": true, "Ev Idx": 11990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911681.711, "dur": 0.102,
    "args": {
      "Python parent id": 9090, "Python id": 9091, "finished": true, "Ev Idx": 11991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911682.119, "dur": 19.300,
    "args": {
      "Python parent id": 9090, "Python id": 9092, "finished": true, "Ev Idx": 11992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911682.245, "dur": 8.558,
    "args": {
      "Python parent id": 9092, "Python id": 9093, "finished": true, "Ev Idx": 11993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911710.107, "dur": 7.604,
    "args": {
      "Python parent id": 9089, "Python id": 9094, "finished": true, "Ev Idx": 11994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911726.835, "dur": 8.880,
    "args": {
      "Python parent id": 9089, "Python id": 9095, "finished": true, "Ev Idx": 11995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911757.535, "dur": 18.578,
    "args": {
      "Python parent id": 12, "Python id": 9096, "finished": true, "Ev Idx": 11996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911757.869, "dur": 12.485,
    "args": {
      "Python parent id": 9096, "Python id": 9097, "finished": true, "Ev Idx": 11997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911758.053, "dur": 0.494,
    "args": {
      "Python parent id": 9097, "Python id": 9098, "finished": true, "Ev Idx": 11998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911758.244, "dur": 0.233,
    "args": {
      "Python parent id": 9098, "Python id": 9099, "finished": true, "Ev Idx": 11999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911758.361, "dur": 0.081,
    "args": {
      "Python parent id": 9099, "Python id": 9100, "finished": true, "Ev Idx": 12000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911758.859, "dur": 9.244,
    "args": {
      "Python parent id": 9097, "Python id": 9101, "finished": true, "Ev Idx": 12001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911758.944, "dur": 0.078,
    "args": {
      "Python parent id": 9101, "Python id": 9102, "finished": true, "Ev Idx": 12002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911759.145, "dur": 0.068,
    "args": {
      "Python parent id": 9101, "Python id": 9103, "finished": true, "Ev Idx": 12003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911759.373, "dur": 0.066,
    "args": {
      "Python parent id": 9101, "Python id": 9104, "finished": true, "Ev Idx": 12004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911759.952, "dur": 0.040,
    "args": {
      "Python parent id": 9101, "Python id": 9105, "finished": true, "Ev Idx": 12005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911760.285, "dur": 0.085,
    "args": {
      "Python parent id": 9101, "Python id": 9106, "finished": true, "Ev Idx": 12006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911760.717, "dur": 7.328,
    "args": {
      "Python parent id": 9101, "Python id": 9107, "finished": true, "Ev Idx": 12007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911760.829, "dur": 0.066,
    "args": {
      "Python parent id": 9107, "Python id": 9108, "finished": true, "Ev Idx": 12008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911761.048, "dur": 0.054,
    "args": {
      "Python parent id": 9107, "Python id": 9109, "finished": true, "Ev Idx": 12009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911761.261, "dur": 0.067,
    "args": {
      "Python parent id": 9107, "Python id": 9110, "finished": true, "Ev Idx": 12010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911761.688, "dur": 0.041,
    "args": {
      "Python parent id": 9107, "Python id": 9111, "finished": true, "Ev Idx": 12011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911762.049, "dur": 5.911,
    "args": {
      "Python parent id": 9107, "Python id": 9112, "finished": true, "Ev Idx": 12012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911762.327, "dur": 5.572,
    "args": {
      "Python parent id": 9112, "Python id": 9113, "finished": true, "Ev Idx": 12013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911762.447, "dur": 3.637,
    "args": {
      "Python parent id": 9113, "Python id": 9114, "finished": true, "Ev Idx": 12014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911762.728, "dur": 3.285,
    "args": {
      "Python parent id": 9114, "Python id": 9115, "finished": true, "Ev Idx": 12015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911762.901, "dur": 0.264,
    "args": {
      "Python parent id": 9115, "Python id": 9116, "finished": true, "Ev Idx": 12016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911763.040, "dur": 0.072,
    "args": {
      "Python parent id": 9116, "Python id": 9117, "finished": true, "Ev Idx": 12017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911763.313, "dur": 2.276,
    "args": {
      "Python parent id": 9115, "Python id": 9118, "finished": true, "Ev Idx": 12018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911763.572, "dur": 1.925,
    "args": {
      "Python parent id": 9118, "Python id": 9119, "finished": true, "Ev Idx": 12019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911763.769, "dur": 1.659,
    "args": {
      "Python parent id": 9119, "Python id": 9120, "finished": true, "Ev Idx": 12020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911763.929, "dur": 1.242,
    "args": {
      "Python parent id": 9120, "Python id": 9121, "finished": true, "Ev Idx": 12021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911764.135, "dur": 0.514,
    "args": {
      "Python parent id": 9121, "Python id": 9122, "finished": true, "Ev Idx": 12022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911764.244, "dur": 0.045,
    "args": {
      "Python parent id": 9122, "Python id": 9123, "finished": true, "Ev Idx": 12023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911764.431, "dur": 0.140,
    "args": {
      "Python parent id": 9122, "Python id": 9124, "finished": true, "Ev Idx": 12024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911765.784, "dur": 0.169,
    "args": {
      "Python parent id": 9115, "Python id": 9125, "finished": true, "Ev Idx": 12025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911766.274, "dur": 0.104,
    "args": {
      "Python parent id": 9113, "Python id": 9126, "finished": true, "Ev Idx": 12026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911766.653, "dur": 1.203,
    "args": {
      "Python parent id": 9113, "Python id": 9127, "finished": true, "Ev Idx": 12027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911766.855, "dur": 0.945,
    "args": {
      "Python parent id": 9127, "Python id": 9128, "finished": true, "Ev Idx": 12028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911767.005, "dur": 0.391,
    "args": {
      "Python parent id": 9128, "Python id": 9129, "finished": true, "Ev Idx": 12029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911767.147, "dur": 0.190,
    "args": {
      "Python parent id": 9129, "Python id": 9130, "finished": true, "Ev Idx": 12030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911767.248, "dur": 0.055,
    "args": {
      "Python parent id": 9130, "Python id": 9131, "finished": true, "Ev Idx": 12031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911767.527, "dur": 0.220,
    "args": {
      "Python parent id": 9128, "Python id": 9132, "finished": true, "Ev Idx": 12032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911768.258, "dur": 0.243,
    "args": {
      "Python parent id": 9097, "Python id": 9133, "finished": true, "Ev Idx": 12033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911769.129, "dur": 1.057,
    "args": {
      "Python parent id": 9097, "Python id": 9134, "finished": true, "Ev Idx": 12034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911769.613, "dur": 0.480,
    "args": {
      "Python parent id": 9134, "Python id": 9135, "finished": true, "Ev Idx": 12035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911770.726, "dur": 5.290,
    "args": {
      "Python parent id": 9096, "Python id": 9136, "finished": true, "Ev Idx": 12036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911776.701, "dur": 8.559,
    "args": {
      "Python parent id": 12, "Python id": 9137, "finished": true, "Ev Idx": 12037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911785.732, "dur": 0.114,
    "args": {
      "Python parent id": 12, "Python id": 9138, "finished": true, "Ev Idx": 12038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911786.487, "dur": 0.209,
    "args": {
      "Python parent id": 12, "Python id": 9139, "finished": true, "Ev Idx": 12039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911787.205, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 9140, "finished": true, "Ev Idx": 12040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911787.723, "dur": 12.734,
    "args": {
      "Python parent id": 12, "Python id": 9141, "finished": true, "Ev Idx": 12041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911801.743, "dur": 16.283,
    "args": {
      "Python parent id": 12, "Python id": 9142, "finished": true, "Ev Idx": 12042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911802.081, "dur": 11.497,
    "args": {
      "Python parent id": 9142, "Python id": 9143, "finished": true, "Ev Idx": 12043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911802.272, "dur": 0.509,
    "args": {
      "Python parent id": 9143, "Python id": 9144, "finished": true, "Ev Idx": 12044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911802.444, "dur": 0.269,
    "args": {
      "Python parent id": 9144, "Python id": 9145, "finished": true, "Ev Idx": 12045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911802.580, "dur": 0.098,
    "args": {
      "Python parent id": 9145, "Python id": 9146, "finished": true, "Ev Idx": 12046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911803.136, "dur": 8.515,
    "args": {
      "Python parent id": 9143, "Python id": 9147, "finished": true, "Ev Idx": 12047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911803.232, "dur": 0.071,
    "args": {
      "Python parent id": 9147, "Python id": 9148, "finished": true, "Ev Idx": 12048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911803.421, "dur": 0.062,
    "args": {
      "Python parent id": 9147, "Python id": 9149, "finished": true, "Ev Idx": 12049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911803.636, "dur": 0.060,
    "args": {
      "Python parent id": 9147, "Python id": 9150, "finished": true, "Ev Idx": 12050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911804.178, "dur": 0.043,
    "args": {
      "Python parent id": 9147, "Python id": 9151, "finished": true, "Ev Idx": 12051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911804.517, "dur": 0.076,
    "args": {
      "Python parent id": 9147, "Python id": 9152, "finished": true, "Ev Idx": 12052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911804.959, "dur": 6.634,
    "args": {
      "Python parent id": 9147, "Python id": 9153, "finished": true, "Ev Idx": 12053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911805.053, "dur": 0.062,
    "args": {
      "Python parent id": 9153, "Python id": 9154, "finished": true, "Ev Idx": 12054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911805.261, "dur": 0.065,
    "args": {
      "Python parent id": 9153, "Python id": 9155, "finished": true, "Ev Idx": 12055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911805.494, "dur": 0.085,
    "args": {
      "Python parent id": 9153, "Python id": 9156, "finished": true, "Ev Idx": 12056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911805.910, "dur": 0.066,
    "args": {
      "Python parent id": 9153, "Python id": 9157, "finished": true, "Ev Idx": 12057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911806.291, "dur": 5.194,
    "args": {
      "Python parent id": 9153, "Python id": 9158, "finished": true, "Ev Idx": 12058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911806.558, "dur": 4.871,
    "args": {
      "Python parent id": 9158, "Python id": 9159, "finished": true, "Ev Idx": 12059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911806.687, "dur": 3.084,
    "args": {
      "Python parent id": 9159, "Python id": 9160, "finished": true, "Ev Idx": 12060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911806.880, "dur": 2.844,
    "args": {
      "Python parent id": 9160, "Python id": 9161, "finished": true, "Ev Idx": 12061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911807.092, "dur": 0.233,
    "args": {
      "Python parent id": 9161, "Python id": 9162, "finished": true, "Ev Idx": 12062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911807.207, "dur": 0.070,
    "args": {
      "Python parent id": 9162, "Python id": 9163, "finished": true, "Ev Idx": 12063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911807.504, "dur": 1.870,
    "args": {
      "Python parent id": 9161, "Python id": 9164, "finished": true, "Ev Idx": 12064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911807.661, "dur": 1.650,
    "args": {
      "Python parent id": 9164, "Python id": 9165, "finished": true, "Ev Idx": 12065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911807.871, "dur": 1.387,
    "args": {
      "Python parent id": 9165, "Python id": 9166, "finished": true, "Ev Idx": 12066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911808.031, "dur": 1.056,
    "args": {
      "Python parent id": 9166, "Python id": 9167, "finished": true, "Ev Idx": 12067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911808.208, "dur": 0.493,
    "args": {
      "Python parent id": 9167, "Python id": 9168, "finished": true, "Ev Idx": 12068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911808.311, "dur": 0.052,
    "args": {
      "Python parent id": 9168, "Python id": 9169, "finished": true, "Ev Idx": 12069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911808.525, "dur": 0.123,
    "args": {
      "Python parent id": 9168, "Python id": 9170, "finished": true, "Ev Idx": 12070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911809.578, "dur": 0.111,
    "args": {
      "Python parent id": 9161, "Python id": 9171, "finished": true, "Ev Idx": 12071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911809.894, "dur": 0.093,
    "args": {
      "Python parent id": 9159, "Python id": 9172, "finished": true, "Ev Idx": 12072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911810.253, "dur": 1.128,
    "args": {
      "Python parent id": 9159, "Python id": 9173, "finished": true, "Ev Idx": 12073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911810.391, "dur": 0.928,
    "args": {
      "Python parent id": 9173, "Python id": 9174, "finished": true, "Ev Idx": 12074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911810.552, "dur": 0.380,
    "args": {
      "Python parent id": 9174, "Python id": 9175, "finished": true, "Ev Idx": 12075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911810.694, "dur": 0.177,
    "args": {
      "Python parent id": 9175, "Python id": 9176, "finished": true, "Ev Idx": 12076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911810.782, "dur": 0.059,
    "args": {
      "Python parent id": 9176, "Python id": 9177, "finished": true, "Ev Idx": 12077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911811.074, "dur": 0.211,
    "args": {
      "Python parent id": 9174, "Python id": 9178, "finished": true, "Ev Idx": 12078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911811.775, "dur": 0.131,
    "args": {
      "Python parent id": 9143, "Python id": 9179, "finished": true, "Ev Idx": 12079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911812.542, "dur": 0.914,
    "args": {
      "Python parent id": 9143, "Python id": 9180, "finished": true, "Ev Idx": 12080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911812.958, "dur": 0.444,
    "args": {
      "Python parent id": 9180, "Python id": 9181, "finished": true, "Ev Idx": 12081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911813.855, "dur": 4.093,
    "args": {
      "Python parent id": 9142, "Python id": 9182, "finished": true, "Ev Idx": 12082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911818.928, "dur": 99.636,
    "args": {
      "Python parent id": 12, "Python id": 9183, "finished": true, "Ev Idx": 12083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911819.291, "dur": 0.469,
    "args": {
      "Python parent id": 9183, "Python id": 9184, "finished": true, "Ev Idx": 12084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911819.584, "dur": 0.134,
    "args": {
      "Python parent id": 9184, "Python id": 9185, "finished": true, "Ev Idx": 12085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911820.090, "dur": 0.395,
    "args": {
      "Python parent id": 9183, "Python id": 9186, "finished": true, "Ev Idx": 12086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911820.706, "dur": 0.285,
    "args": {
      "Python parent id": 9183, "Python id": 9187, "finished": true, "Ev Idx": 12087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911821.197, "dur": 1.498,
    "args": {
      "Python parent id": 9183, "Python id": 9188, "finished": true, "Ev Idx": 12088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911823.355, "dur": 95.106,
    "args": {
      "Python parent id": 9183, "Python id": 9189, "finished": true, "Ev Idx": 12089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911824.345, "dur": 93.853,
    "args": {
      "Python parent id": 9189, "Python id": 9190, "finished": true, "Ev Idx": 12090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911824.572, "dur": 93.561,
    "args": {
      "Python parent id": 9190, "Python id": 9191, "finished": true, "Ev Idx": 12091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911831.242, "dur": 85.389,
    "args": {
      "Python parent id": 9191, "Python id": 9192, "finished": true, "Ev Idx": 12092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911831.640, "dur": 84.910,
    "args": {
      "Python parent id": 9192, "Python id": 9193, "finished": true, "Ev Idx": 12093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911832.174, "dur": 84.210,
    "args": {
      "Python parent id": 9193, "Python id": 9194, "finished": true, "Ev Idx": 12094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911832.777, "dur": 0.423,
    "args": {
      "Python parent id": 9194, "Python id": 9195, "finished": true, "Ev Idx": 12095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911833.070, "dur": 0.041,
    "args": {
      "Python parent id": 9195, "Python id": 9196, "finished": true, "Ev Idx": 12096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911833.513, "dur": 0.381,
    "args": {
      "Python parent id": 9194, "Python id": 9197, "finished": true, "Ev Idx": 12097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911833.806, "dur": 0.033,
    "args": {
      "Python parent id": 9197, "Python id": 9198, "finished": true, "Ev Idx": 12098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911834.248, "dur": 81.963,
    "args": {
      "Python parent id": 9194, "Python id": 9199, "finished": true, "Ev Idx": 12099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911835.029, "dur": 79.845,
    "args": {
      "Python parent id": 9199, "Python id": 9200, "finished": true, "Ev Idx": 12100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911843.637, "dur": 20.418,
    "args": {
      "Python parent id": 9200, "Python id": 9201, "finished": true, "Ev Idx": 12101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911843.805, "dur": 0.142,
    "args": {
      "Python parent id": 9201, "Python id": 9202, "finished": true, "Ev Idx": 12102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740911844.261, "dur": 19.584,
    "args": {
      "Python parent id": 9201, "Python id": 9203, "finished": true, "Ev Idx": 12103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911844.390, "dur": 9.200,
    "args": {
      "Python parent id": 9203, "Python id": 9204, "finished": true, "Ev Idx": 12104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911872.329, "dur": 7.452,
    "args": {
      "Python parent id": 9200, "Python id": 9205, "finished": true, "Ev Idx": 12105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740911889.313, "dur": 8.760,
    "args": {
      "Python parent id": 9200, "Python id": 9206, "finished": true, "Ev Idx": 12106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911919.840, "dur": 17.926,
    "args": {
      "Python parent id": 12, "Python id": 9207, "finished": true, "Ev Idx": 12107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911920.245, "dur": 13.489,
    "args": {
      "Python parent id": 9207, "Python id": 9208, "finished": true, "Ev Idx": 12108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911920.446, "dur": 0.575,
    "args": {
      "Python parent id": 9208, "Python id": 9209, "finished": true, "Ev Idx": 12109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911920.642, "dur": 0.291,
    "args": {
      "Python parent id": 9209, "Python id": 9210, "finished": true, "Ev Idx": 12110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911920.795, "dur": 0.101,
    "args": {
      "Python parent id": 9210, "Python id": 9211, "finished": true, "Ev Idx": 12111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911921.463, "dur": 10.067,
    "args": {
      "Python parent id": 9208, "Python id": 9212, "finished": true, "Ev Idx": 12112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911921.600, "dur": 0.069,
    "args": {
      "Python parent id": 9212, "Python id": 9213, "finished": true, "Ev Idx": 12113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911921.809, "dur": 0.063,
    "args": {
      "Python parent id": 9212, "Python id": 9214, "finished": true, "Ev Idx": 12114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911922.013, "dur": 0.065,
    "args": {
      "Python parent id": 9212, "Python id": 9215, "finished": true, "Ev Idx": 12115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911922.496, "dur": 0.040,
    "args": {
      "Python parent id": 9212, "Python id": 9216, "finished": true, "Ev Idx": 12116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911922.798, "dur": 0.085,
    "args": {
      "Python parent id": 9212, "Python id": 9217, "finished": true, "Ev Idx": 12117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911923.270, "dur": 8.184,
    "args": {
      "Python parent id": 9212, "Python id": 9218, "finished": true, "Ev Idx": 12118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911923.374, "dur": 0.070,
    "args": {
      "Python parent id": 9218, "Python id": 9219, "finished": true, "Ev Idx": 12119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911923.593, "dur": 0.068,
    "args": {
      "Python parent id": 9218, "Python id": 9220, "finished": true, "Ev Idx": 12120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911923.833, "dur": 0.070,
    "args": {
      "Python parent id": 9218, "Python id": 9221, "finished": true, "Ev Idx": 12121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911924.262, "dur": 0.059,
    "args": {
      "Python parent id": 9218, "Python id": 9222, "finished": true, "Ev Idx": 12122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911924.661, "dur": 6.731,
    "args": {
      "Python parent id": 9218, "Python id": 9223, "finished": true, "Ev Idx": 12123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911924.975, "dur": 6.359,
    "args": {
      "Python parent id": 9223, "Python id": 9224, "finished": true, "Ev Idx": 12124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911925.122, "dur": 4.232,
    "args": {
      "Python parent id": 9224, "Python id": 9225, "finished": true, "Ev Idx": 12125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911925.321, "dur": 3.947,
    "args": {
      "Python parent id": 9225, "Python id": 9226, "finished": true, "Ev Idx": 12126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911925.461, "dur": 0.281,
    "args": {
      "Python parent id": 9226, "Python id": 9227, "finished": true, "Ev Idx": 12127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911925.611, "dur": 0.067,
    "args": {
      "Python parent id": 9227, "Python id": 9228, "finished": true, "Ev Idx": 12128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911925.889, "dur": 2.937,
    "args": {
      "Python parent id": 9226, "Python id": 9229, "finished": true, "Ev Idx": 12129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911926.135, "dur": 2.555,
    "args": {
      "Python parent id": 9229, "Python id": 9230, "finished": true, "Ev Idx": 12130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911926.387, "dur": 2.254,
    "args": {
      "Python parent id": 9230, "Python id": 9231, "finished": true, "Ev Idx": 12131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911926.542, "dur": 1.837,
    "args": {
      "Python parent id": 9231, "Python id": 9232, "finished": true, "Ev Idx": 12132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911926.802, "dur": 0.975,
    "args": {
      "Python parent id": 9232, "Python id": 9233, "finished": true, "Ev Idx": 12133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911926.887, "dur": 0.376,
    "args": {
      "Python parent id": 9233, "Python id": 9234, "finished": true, "Ev Idx": 12134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911927.543, "dur": 0.152,
    "args": {
      "Python parent id": 9233, "Python id": 9235, "finished": true, "Ev Idx": 12135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911929.046, "dur": 0.173,
    "args": {
      "Python parent id": 9226, "Python id": 9236, "finished": true, "Ev Idx": 12136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911929.569, "dur": 0.119,
    "args": {
      "Python parent id": 9224, "Python id": 9237, "finished": true, "Ev Idx": 12137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.006, "dur": 1.262,
    "args": {
      "Python parent id": 9224, "Python id": 9238, "finished": true, "Ev Idx": 12138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.174, "dur": 1.043,
    "args": {
      "Python parent id": 9238, "Python id": 9239, "finished": true, "Ev Idx": 12139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.324, "dur": 0.405,
    "args": {
      "Python parent id": 9239, "Python id": 9240, "finished": true, "Ev Idx": 12140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.463, "dur": 0.208,
    "args": {
      "Python parent id": 9240, "Python id": 9241, "finished": true, "Ev Idx": 12141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.556, "dur": 0.083,
    "args": {
      "Python parent id": 9241, "Python id": 9242, "finished": true, "Ev Idx": 12142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911930.872, "dur": 0.281,
    "args": {
      "Python parent id": 9239, "Python id": 9243, "finished": true, "Ev Idx": 12143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911931.678, "dur": 0.245,
    "args": {
      "Python parent id": 9208, "Python id": 9244, "finished": true, "Ev Idx": 12144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911932.553, "dur": 1.017,
    "args": {
      "Python parent id": 9208, "Python id": 9245, "finished": true, "Ev Idx": 12145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911933.017, "dur": 0.501,
    "args": {
      "Python parent id": 9245, "Python id": 9246, "finished": true, "Ev Idx": 12146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911934.135, "dur": 3.554,
    "args": {
      "Python parent id": 9207, "Python id": 9247, "finished": true, "Ev Idx": 12147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911938.340, "dur": 8.552,
    "args": {
      "Python parent id": 12, "Python id": 9248, "finished": true, "Ev Idx": 12148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740911947.337, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 9249, "finished": true, "Ev Idx": 12149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911948.038, "dur": 0.247,
    "args": {
      "Python parent id": 12, "Python id": 9250, "finished": true, "Ev Idx": 12150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911948.861, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 9251, "finished": true, "Ev Idx": 12151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740911949.349, "dur": 10.456,
    "args": {
      "Python parent id": 12, "Python id": 9252, "finished": true, "Ev Idx": 12152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740911961.323, "dur": 15.345,
    "args": {
      "Python parent id": 12, "Python id": 9253, "finished": true, "Ev Idx": 12153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740911961.632, "dur": 11.571,
    "args": {
      "Python parent id": 9253, "Python id": 9254, "finished": true, "Ev Idx": 12154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911961.843, "dur": 0.474,
    "args": {
      "Python parent id": 9254, "Python id": 9255, "finished": true, "Ev Idx": 12155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911962.007, "dur": 0.234,
    "args": {
      "Python parent id": 9255, "Python id": 9256, "finished": true, "Ev Idx": 12156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911962.094, "dur": 0.110,
    "args": {
      "Python parent id": 9256, "Python id": 9257, "finished": true, "Ev Idx": 12157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911962.613, "dur": 8.749,
    "args": {
      "Python parent id": 9254, "Python id": 9258, "finished": true, "Ev Idx": 12158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911962.745, "dur": 0.069,
    "args": {
      "Python parent id": 9258, "Python id": 9259, "finished": true, "Ev Idx": 12159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911962.953, "dur": 0.060,
    "args": {
      "Python parent id": 9258, "Python id": 9260, "finished": true, "Ev Idx": 12160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911963.159, "dur": 0.063,
    "args": {
      "Python parent id": 9258, "Python id": 9261, "finished": true, "Ev Idx": 12161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911963.669, "dur": 0.062,
    "args": {
      "Python parent id": 9258, "Python id": 9262, "finished": true, "Ev Idx": 12162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911964.015, "dur": 0.082,
    "args": {
      "Python parent id": 9258, "Python id": 9263, "finished": true, "Ev Idx": 12163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911964.437, "dur": 6.827,
    "args": {
      "Python parent id": 9258, "Python id": 9264, "finished": true, "Ev Idx": 12164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911964.549, "dur": 0.065,
    "args": {
      "Python parent id": 9264, "Python id": 9265, "finished": true, "Ev Idx": 12165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911964.791, "dur": 0.059,
    "args": {
      "Python parent id": 9264, "Python id": 9266, "finished": true, "Ev Idx": 12166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911965.051, "dur": 0.069,
    "args": {
      "Python parent id": 9264, "Python id": 9267, "finished": true, "Ev Idx": 12167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740911965.469, "dur": 0.062,
    "args": {
      "Python parent id": 9264, "Python id": 9268, "finished": true, "Ev Idx": 12168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740911965.842, "dur": 5.338,
    "args": {
      "Python parent id": 9264, "Python id": 9269, "finished": true, "Ev Idx": 12169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.087, "dur": 5.036,
    "args": {
      "Python parent id": 9269, "Python id": 9270, "finished": true, "Ev Idx": 12170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.213, "dur": 3.098,
    "args": {
      "Python parent id": 9270, "Python id": 9271, "finished": true, "Ev Idx": 12171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.401, "dur": 2.862,
    "args": {
      "Python parent id": 9271, "Python id": 9272, "finished": true, "Ev Idx": 12172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.520, "dur": 0.246,
    "args": {
      "Python parent id": 9272, "Python id": 9273, "finished": true, "Ev Idx": 12173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.649, "dur": 0.065,
    "args": {
      "Python parent id": 9273, "Python id": 9274, "finished": true, "Ev Idx": 12174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740911966.951, "dur": 1.981,
    "args": {
      "Python parent id": 9272, "Python id": 9275, "finished": true, "Ev Idx": 12175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.151, "dur": 1.721,
    "args": {
      "Python parent id": 9275, "Python id": 9276, "finished": true, "Ev Idx": 12176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.367, "dur": 1.457,
    "args": {
      "Python parent id": 9276, "Python id": 9277, "finished": true, "Ev Idx": 12177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.521, "dur": 1.118,
    "args": {
      "Python parent id": 9277, "Python id": 9278, "finished": true, "Ev Idx": 12178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.706, "dur": 0.486,
    "args": {
      "Python parent id": 9278, "Python id": 9279, "finished": true, "Ev Idx": 12179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.787, "dur": 0.051,
    "args": {
      "Python parent id": 9279, "Python id": 9280, "finished": true, "Ev Idx": 12180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740911967.996, "dur": 0.124,
    "args": {
      "Python parent id": 9279, "Python id": 9281, "finished": true, "Ev Idx": 12181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740911969.114, "dur": 0.114,
    "args": {
      "Python parent id": 9272, "Python id": 9282, "finished": true, "Ev Idx": 12182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740911969.470, "dur": 0.116,
    "args": {
      "Python parent id": 9270, "Python id": 9283, "finished": true, "Ev Idx": 12183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740911969.879, "dur": 1.200,
    "args": {
      "Python parent id": 9270, "Python id": 9284, "finished": true, "Ev Idx": 12184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740911970.048, "dur": 0.987,
    "args": {
      "Python parent id": 9284, "Python id": 9285, "finished": true, "Ev Idx": 12185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740911970.203, "dur": 0.411,
    "args": {
      "Python parent id": 9285, "Python id": 9286, "finished": true, "Ev Idx": 12186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740911970.377, "dur": 0.177,
    "args": {
      "Python parent id": 9286, "Python id": 9287, "finished": true, "Ev Idx": 12187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740911970.459, "dur": 0.060,
    "args": {
      "Python parent id": 9287, "Python id": 9288, "finished": true, "Ev Idx": 12188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740911970.746, "dur": 0.256,
    "args": {
      "Python parent id": 9285, "Python id": 9289, "finished": true, "Ev Idx": 12189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740911971.511, "dur": 0.152,
    "args": {
      "Python parent id": 9254, "Python id": 9290, "finished": true, "Ev Idx": 12190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740911972.225, "dur": 0.823,
    "args": {
      "Python parent id": 9254, "Python id": 9291, "finished": true, "Ev Idx": 12191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740911972.608, "dur": 0.386,
    "args": {
      "Python parent id": 9291, "Python id": 9292, "finished": true, "Ev Idx": 12192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740911973.451, "dur": 3.159,
    "args": {
      "Python parent id": 9253, "Python id": 9293, "finished": true, "Ev Idx": 12193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740911977.633, "dur": 91.369,
    "args": {
      "Python parent id": 12, "Python id": 9294, "finished": true, "Ev Idx": 12194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911977.938, "dur": 0.469,
    "args": {
      "Python parent id": 9294, "Python id": 9295, "finished": true, "Ev Idx": 12195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911978.236, "dur": 0.083,
    "args": {
      "Python parent id": 9295, "Python id": 9296, "finished": true, "Ev Idx": 12196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911978.728, "dur": 0.506,
    "args": {
      "Python parent id": 9294, "Python id": 9297, "finished": true, "Ev Idx": 12197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740911979.487, "dur": 0.288,
    "args": {
      "Python parent id": 9294, "Python id": 9298, "finished": true, "Ev Idx": 12198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740911980.007, "dur": 1.563,
    "args": {
      "Python parent id": 9294, "Python id": 9299, "finished": true, "Ev Idx": 12199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740911982.125, "dur": 86.768,
    "args": {
      "Python parent id": 9294, "Python id": 9300, "finished": true, "Ev Idx": 12200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740911982.960, "dur": 85.772,
    "args": {
      "Python parent id": 9300, "Python id": 9301, "finished": true, "Ev Idx": 12201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740911983.300, "dur": 85.351,
    "args": {
      "Python parent id": 9301, "Python id": 9302, "finished": true, "Ev Idx": 12202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740911988.746, "dur": 78.593,
    "args": {
      "Python parent id": 9302, "Python id": 9303, "finished": true, "Ev Idx": 12203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740911989.152, "dur": 78.131,
    "args": {
      "Python parent id": 9303, "Python id": 9304, "finished": true, "Ev Idx": 12204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740911989.852, "dur": 77.300,
    "args": {
      "Python parent id": 9304, "Python id": 9305, "finished": true, "Ev Idx": 12205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911990.478, "dur": 0.412,
    "args": {
      "Python parent id": 9305, "Python id": 9306, "finished": true, "Ev Idx": 12206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911990.758, "dur": 0.054,
    "args": {
      "Python parent id": 9306, "Python id": 9307, "finished": true, "Ev Idx": 12207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740911991.246, "dur": 0.386,
    "args": {
      "Python parent id": 9305, "Python id": 9308, "finished": true, "Ev Idx": 12208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740911991.503, "dur": 0.078,
    "args": {
      "Python parent id": 9308, "Python id": 9309, "finished": true, "Ev Idx": 12209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911992.034, "dur": 75.018,
    "args": {
      "Python parent id": 9305, "Python id": 9310, "finished": true, "Ev Idx": 12210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740911992.683, "dur": 73.100,
    "args": {
      "Python parent id": 9310, "Python id": 9311, "finished": true, "Ev Idx": 12211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740911999.847, "dur": 17.875,
    "args": {
      "Python parent id": 9311, "Python id": 9312, "finished": true, "Ev Idx": 12212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740911999.955, "dur": 0.106,
    "args": {
      "Python parent id": 9312, "Python id": 9313, "finished": true, "Ev Idx": 12213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912000.355, "dur": 17.229,
    "args": {
      "Python parent id": 9312, "Python id": 9314, "finished": true, "Ev Idx": 12214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912000.518, "dur": 7.144,
    "args": {
      "Python parent id": 9314, "Python id": 9315, "finished": true, "Ev Idx": 12215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912025.769, "dur": 7.425,
    "args": {
      "Python parent id": 9311, "Python id": 9316, "finished": true, "Ev Idx": 12216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912041.751, "dur": 7.994,
    "args": {
      "Python parent id": 9311, "Python id": 9317, "finished": true, "Ev Idx": 12217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912070.288, "dur": 17.106,
    "args": {
      "Python parent id": 12, "Python id": 9318, "finished": true, "Ev Idx": 12218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912070.671, "dur": 12.807,
    "args": {
      "Python parent id": 9318, "Python id": 9319, "finished": true, "Ev Idx": 12219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912070.900, "dur": 0.502,
    "args": {
      "Python parent id": 9319, "Python id": 9320, "finished": true, "Ev Idx": 12220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912071.109, "dur": 0.227,
    "args": {
      "Python parent id": 9320, "Python id": 9321, "finished": true, "Ev Idx": 12221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912071.225, "dur": 0.074,
    "args": {
      "Python parent id": 9321, "Python id": 9322, "finished": true, "Ev Idx": 12222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912071.844, "dur": 9.543,
    "args": {
      "Python parent id": 9319, "Python id": 9323, "finished": true, "Ev Idx": 12223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912071.959, "dur": 0.078,
    "args": {
      "Python parent id": 9323, "Python id": 9324, "finished": true, "Ev Idx": 12224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912072.168, "dur": 0.059,
    "args": {
      "Python parent id": 9323, "Python id": 9325, "finished": true, "Ev Idx": 12225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912072.403, "dur": 0.068,
    "args": {
      "Python parent id": 9323, "Python id": 9326, "finished": true, "Ev Idx": 12226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912072.931, "dur": 0.048,
    "args": {
      "Python parent id": 9323, "Python id": 9327, "finished": true, "Ev Idx": 12227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912073.279, "dur": 0.094,
    "args": {
      "Python parent id": 9323, "Python id": 9328, "finished": true, "Ev Idx": 12228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912073.706, "dur": 7.589,
    "args": {
      "Python parent id": 9323, "Python id": 9329, "finished": true, "Ev Idx": 12229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912073.821, "dur": 0.062,
    "args": {
      "Python parent id": 9329, "Python id": 9330, "finished": true, "Ev Idx": 12230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912074.053, "dur": 0.064,
    "args": {
      "Python parent id": 9329, "Python id": 9331, "finished": true, "Ev Idx": 12231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912074.302, "dur": 0.073,
    "args": {
      "Python parent id": 9329, "Python id": 9332, "finished": true, "Ev Idx": 12232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912074.756, "dur": 0.055,
    "args": {
      "Python parent id": 9329, "Python id": 9333, "finished": true, "Ev Idx": 12233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912075.147, "dur": 6.064,
    "args": {
      "Python parent id": 9329, "Python id": 9334, "finished": true, "Ev Idx": 12234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912075.441, "dur": 5.704,
    "args": {
      "Python parent id": 9334, "Python id": 9335, "finished": true, "Ev Idx": 12235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912075.569, "dur": 3.546,
    "args": {
      "Python parent id": 9335, "Python id": 9336, "finished": true, "Ev Idx": 12236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912075.795, "dur": 3.271,
    "args": {
      "Python parent id": 9336, "Python id": 9337, "finished": true, "Ev Idx": 12237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912075.936, "dur": 0.270,
    "args": {
      "Python parent id": 9337, "Python id": 9338, "finished": true, "Ev Idx": 12238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912076.065, "dur": 0.067,
    "args": {
      "Python parent id": 9338, "Python id": 9339, "finished": true, "Ev Idx": 12239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912076.399, "dur": 2.229,
    "args": {
      "Python parent id": 9337, "Python id": 9340, "finished": true, "Ev Idx": 12240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912076.700, "dur": 1.852,
    "args": {
      "Python parent id": 9340, "Python id": 9341, "finished": true, "Ev Idx": 12241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912076.897, "dur": 1.607,
    "args": {
      "Python parent id": 9341, "Python id": 9342, "finished": true, "Ev Idx": 12242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912077.084, "dur": 1.220,
    "args": {
      "Python parent id": 9342, "Python id": 9343, "finished": true, "Ev Idx": 12243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912077.264, "dur": 0.555,
    "args": {
      "Python parent id": 9343, "Python id": 9344, "finished": true, "Ev Idx": 12244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912077.379, "dur": 0.041,
    "args": {
      "Python parent id": 9344, "Python id": 9345, "finished": true, "Ev Idx": 12245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912077.627, "dur": 0.157,
    "args": {
      "Python parent id": 9344, "Python id": 9346, "finished": true, "Ev Idx": 12246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912078.832, "dur": 0.169,
    "args": {
      "Python parent id": 9337, "Python id": 9347, "finished": true, "Ev Idx": 12247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912079.312, "dur": 0.119,
    "args": {
      "Python parent id": 9335, "Python id": 9348, "finished": true, "Ev Idx": 12248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912079.681, "dur": 1.404,
    "args": {
      "Python parent id": 9335, "Python id": 9349, "finished": true, "Ev Idx": 12249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912079.891, "dur": 1.149,
    "args": {
      "Python parent id": 9349, "Python id": 9350, "finished": true, "Ev Idx": 12250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912080.056, "dur": 0.400,
    "args": {
      "Python parent id": 9350, "Python id": 9351, "finished": true, "Ev Idx": 12251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912080.204, "dur": 0.193,
    "args": {
      "Python parent id": 9351, "Python id": 9352, "finished": true, "Ev Idx": 12252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912080.311, "dur": 0.055,
    "args": {
      "Python parent id": 9352, "Python id": 9353, "finished": true, "Ev Idx": 12253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912080.591, "dur": 0.405,
    "args": {
      "Python parent id": 9350, "Python id": 9354, "finished": true, "Ev Idx": 12254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912081.529, "dur": 0.205,
    "args": {
      "Python parent id": 9319, "Python id": 9355, "finished": true, "Ev Idx": 12255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912082.400, "dur": 0.941,
    "args": {
      "Python parent id": 9319, "Python id": 9356, "finished": true, "Ev Idx": 12256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912082.850, "dur": 0.437,
    "args": {
      "Python parent id": 9356, "Python id": 9357, "finished": true, "Ev Idx": 12257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912083.749, "dur": 3.590,
    "args": {
      "Python parent id": 9318, "Python id": 9358, "finished": true, "Ev Idx": 12258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912087.977, "dur": 7.402,
    "args": {
      "Python parent id": 12, "Python id": 9359, "finished": true, "Ev Idx": 12259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912095.877, "dur": 0.095,
    "args": {
      "Python parent id": 12, "Python id": 9360, "finished": true, "Ev Idx": 12260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912096.541, "dur": 0.223,
    "args": {
      "Python parent id": 12, "Python id": 9361, "finished": true, "Ev Idx": 12261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912097.319, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 9362, "finished": true, "Ev Idx": 12262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912097.807, "dur": 10.423,
    "args": {
      "Python parent id": 12, "Python id": 9363, "finished": true, "Ev Idx": 12263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912109.577, "dur": 15.224,
    "args": {
      "Python parent id": 12, "Python id": 9364, "finished": true, "Ev Idx": 12264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912109.839, "dur": 11.507,
    "args": {
      "Python parent id": 9364, "Python id": 9365, "finished": true, "Ev Idx": 12265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912110.000, "dur": 0.640,
    "args": {
      "Python parent id": 9365, "Python id": 9366, "finished": true, "Ev Idx": 12266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912110.249, "dur": 0.259,
    "args": {
      "Python parent id": 9366, "Python id": 9367, "finished": true, "Ev Idx": 12267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912110.347, "dur": 0.125,
    "args": {
      "Python parent id": 9367, "Python id": 9368, "finished": true, "Ev Idx": 12268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912110.919, "dur": 8.457,
    "args": {
      "Python parent id": 9365, "Python id": 9369, "finished": true, "Ev Idx": 12269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912111.008, "dur": 0.064,
    "args": {
      "Python parent id": 9369, "Python id": 9370, "finished": true, "Ev Idx": 12270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912111.192, "dur": 0.064,
    "args": {
      "Python parent id": 9369, "Python id": 9371, "finished": true, "Ev Idx": 12271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912111.415, "dur": 0.064,
    "args": {
      "Python parent id": 9369, "Python id": 9372, "finished": true, "Ev Idx": 12272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912111.827, "dur": 0.095,
    "args": {
      "Python parent id": 9369, "Python id": 9373, "finished": true, "Ev Idx": 12273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912112.195, "dur": 0.082,
    "args": {
      "Python parent id": 9369, "Python id": 9374, "finished": true, "Ev Idx": 12274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912112.619, "dur": 6.692,
    "args": {
      "Python parent id": 9369, "Python id": 9375, "finished": true, "Ev Idx": 12275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912112.771, "dur": 0.061,
    "args": {
      "Python parent id": 9375, "Python id": 9376, "finished": true, "Ev Idx": 12276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912112.982, "dur": 0.059,
    "args": {
      "Python parent id": 9375, "Python id": 9377, "finished": true, "Ev Idx": 12277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912113.233, "dur": 0.070,
    "args": {
      "Python parent id": 9375, "Python id": 9378, "finished": true, "Ev Idx": 12278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912113.631, "dur": 0.076,
    "args": {
      "Python parent id": 9375, "Python id": 9379, "finished": true, "Ev Idx": 12279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.033, "dur": 5.184,
    "args": {
      "Python parent id": 9375, "Python id": 9380, "finished": true, "Ev Idx": 12280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.283, "dur": 4.874,
    "args": {
      "Python parent id": 9380, "Python id": 9381, "finished": true, "Ev Idx": 12281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.407, "dur": 3.013,
    "args": {
      "Python parent id": 9381, "Python id": 9382, "finished": true, "Ev Idx": 12282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.623, "dur": 2.750,
    "args": {
      "Python parent id": 9382, "Python id": 9383, "finished": true, "Ev Idx": 12283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.824, "dur": 0.235,
    "args": {
      "Python parent id": 9383, "Python id": 9384, "finished": true, "Ev Idx": 12284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912114.941, "dur": 0.079,
    "args": {
      "Python parent id": 9384, "Python id": 9385, "finished": true, "Ev Idx": 12285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.213, "dur": 1.835,
    "args": {
      "Python parent id": 9383, "Python id": 9386, "finished": true, "Ev Idx": 12286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.390, "dur": 1.597,
    "args": {
      "Python parent id": 9386, "Python id": 9387, "finished": true, "Ev Idx": 12287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.570, "dur": 1.373,
    "args": {
      "Python parent id": 9387, "Python id": 9388, "finished": true, "Ev Idx": 12288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.705, "dur": 1.022,
    "args": {
      "Python parent id": 9388, "Python id": 9389, "finished": true, "Ev Idx": 12289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.864, "dur": 0.440,
    "args": {
      "Python parent id": 9389, "Python id": 9390, "finished": true, "Ev Idx": 12290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912115.951, "dur": 0.052,
    "args": {
      "Python parent id": 9390, "Python id": 9391, "finished": true, "Ev Idx": 12291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912116.140, "dur": 0.128,
    "args": {
      "Python parent id": 9390, "Python id": 9392, "finished": true, "Ev Idx": 12292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912117.234, "dur": 0.105,
    "args": {
      "Python parent id": 9383, "Python id": 9393, "finished": true, "Ev Idx": 12293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912117.583, "dur": 0.097,
    "args": {
      "Python parent id": 9381, "Python id": 9394, "finished": true, "Ev Idx": 12294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912117.951, "dur": 1.161,
    "args": {
      "Python parent id": 9381, "Python id": 9395, "finished": true, "Ev Idx": 12295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912118.090, "dur": 0.977,
    "args": {
      "Python parent id": 9395, "Python id": 9396, "finished": true, "Ev Idx": 12296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912118.268, "dur": 0.367,
    "args": {
      "Python parent id": 9396, "Python id": 9397, "finished": true, "Ev Idx": 12297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912118.399, "dur": 0.174,
    "args": {
      "Python parent id": 9397, "Python id": 9398, "finished": true, "Ev Idx": 12298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912118.489, "dur": 0.052,
    "args": {
      "Python parent id": 9398, "Python id": 9399, "finished": true, "Ev Idx": 12299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912118.767, "dur": 0.271,
    "args": {
      "Python parent id": 9396, "Python id": 9400, "finished": true, "Ev Idx": 12300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912119.521, "dur": 0.151,
    "args": {
      "Python parent id": 9365, "Python id": 9401, "finished": true, "Ev Idx": 12301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912120.293, "dur": 0.931,
    "args": {
      "Python parent id": 9365, "Python id": 9402, "finished": true, "Ev Idx": 12302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912120.753, "dur": 0.417,
    "args": {
      "Python parent id": 9402, "Python id": 9403, "finished": true, "Ev Idx": 12303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912121.632, "dur": 3.111,
    "args": {
      "Python parent id": 9364, "Python id": 9404, "finished": true, "Ev Idx": 12304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912125.824, "dur": 92.575,
    "args": {
      "Python parent id": 12, "Python id": 9405, "finished": true, "Ev Idx": 12305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912126.159, "dur": 0.452,
    "args": {
      "Python parent id": 9405, "Python id": 9406, "finished": true, "Ev Idx": 12306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912126.460, "dur": 0.097,
    "args": {
      "Python parent id": 9406, "Python id": 9407, "finished": true, "Ev Idx": 12307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912126.896, "dur": 0.428,
    "args": {
      "Python parent id": 9405, "Python id": 9408, "finished": true, "Ev Idx": 12308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912127.597, "dur": 0.267,
    "args": {
      "Python parent id": 9405, "Python id": 9409, "finished": true, "Ev Idx": 12309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912128.083, "dur": 1.603,
    "args": {
      "Python parent id": 9405, "Python id": 9410, "finished": true, "Ev Idx": 12310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912130.306, "dur": 87.959,
    "args": {
      "Python parent id": 9405, "Python id": 9411, "finished": true, "Ev Idx": 12311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912131.141, "dur": 86.892,
    "args": {
      "Python parent id": 9411, "Python id": 9412, "finished": true, "Ev Idx": 12312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912131.421, "dur": 86.546,
    "args": {
      "Python parent id": 9412, "Python id": 9413, "finished": true, "Ev Idx": 12313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912136.508, "dur": 80.196,
    "args": {
      "Python parent id": 9413, "Python id": 9414, "finished": true, "Ev Idx": 12314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912136.797, "dur": 79.826,
    "args": {
      "Python parent id": 9414, "Python id": 9415, "finished": true, "Ev Idx": 12315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912137.347, "dur": 79.170,
    "args": {
      "Python parent id": 9415, "Python id": 9416, "finished": true, "Ev Idx": 12316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912137.905, "dur": 0.381,
    "args": {
      "Python parent id": 9416, "Python id": 9417, "finished": true, "Ev Idx": 12317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912138.184, "dur": 0.042,
    "args": {
      "Python parent id": 9417, "Python id": 9418, "finished": true, "Ev Idx": 12318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912138.586, "dur": 0.347,
    "args": {
      "Python parent id": 9416, "Python id": 9419, "finished": true, "Ev Idx": 12319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912138.842, "dur": 0.034,
    "args": {
      "Python parent id": 9419, "Python id": 9420, "finished": true, "Ev Idx": 12320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912139.321, "dur": 77.049,
    "args": {
      "Python parent id": 9416, "Python id": 9421, "finished": true, "Ev Idx": 12321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912140.046, "dur": 74.924,
    "args": {
      "Python parent id": 9421, "Python id": 9422, "finished": true, "Ev Idx": 12322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912147.308, "dur": 20.540,
    "args": {
      "Python parent id": 9422, "Python id": 9423, "finished": true, "Ev Idx": 12323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912147.418, "dur": 0.136,
    "args": {
      "Python parent id": 9423, "Python id": 9424, "finished": true, "Ev Idx": 12324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912147.866, "dur": 19.862,
    "args": {
      "Python parent id": 9423, "Python id": 9425, "finished": true, "Ev Idx": 12325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912148.000, "dur": 10.186,
    "args": {
      "Python parent id": 9425, "Python id": 9426, "finished": true, "Ev Idx": 12326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912175.547, "dur": 7.083,
    "args": {
      "Python parent id": 9422, "Python id": 9427, "finished": true, "Ev Idx": 12327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912190.768, "dur": 8.713,
    "args": {
      "Python parent id": 9422, "Python id": 9428, "finished": true, "Ev Idx": 12328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912219.577, "dur": 16.767,
    "args": {
      "Python parent id": 12, "Python id": 9429, "finished": true, "Ev Idx": 12329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912219.953, "dur": 12.105,
    "args": {
      "Python parent id": 9429, "Python id": 9430, "finished": true, "Ev Idx": 12330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912220.185, "dur": 0.518,
    "args": {
      "Python parent id": 9430, "Python id": 9431, "finished": true, "Ev Idx": 12331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912220.417, "dur": 0.186,
    "args": {
      "Python parent id": 9431, "Python id": 9432, "finished": true, "Ev Idx": 12332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912220.502, "dur": 0.067,
    "args": {
      "Python parent id": 9432, "Python id": 9433, "finished": true, "Ev Idx": 12333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912221.123, "dur": 8.818,
    "args": {
      "Python parent id": 9430, "Python id": 9434, "finished": true, "Ev Idx": 12334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912221.204, "dur": 0.068,
    "args": {
      "Python parent id": 9434, "Python id": 9435, "finished": true, "Ev Idx": 12335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912221.390, "dur": 0.059,
    "args": {
      "Python parent id": 9434, "Python id": 9436, "finished": true, "Ev Idx": 12336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912221.595, "dur": 0.065,
    "args": {
      "Python parent id": 9434, "Python id": 9437, "finished": true, "Ev Idx": 12337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912222.115, "dur": 0.067,
    "args": {
      "Python parent id": 9434, "Python id": 9438, "finished": true, "Ev Idx": 12338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912222.474, "dur": 0.079,
    "args": {
      "Python parent id": 9434, "Python id": 9439, "finished": true, "Ev Idx": 12339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912222.925, "dur": 6.947,
    "args": {
      "Python parent id": 9434, "Python id": 9440, "finished": true, "Ev Idx": 12340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912223.042, "dur": 0.064,
    "args": {
      "Python parent id": 9440, "Python id": 9441, "finished": true, "Ev Idx": 12341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912223.262, "dur": 0.061,
    "args": {
      "Python parent id": 9440, "Python id": 9442, "finished": true, "Ev Idx": 12342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912223.508, "dur": 0.070,
    "args": {
      "Python parent id": 9440, "Python id": 9443, "finished": true, "Ev Idx": 12343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912223.909, "dur": 0.068,
    "args": {
      "Python parent id": 9440, "Python id": 9444, "finished": true, "Ev Idx": 12344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912224.307, "dur": 5.506,
    "args": {
      "Python parent id": 9440, "Python id": 9445, "finished": true, "Ev Idx": 12345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912224.606, "dur": 5.145,
    "args": {
      "Python parent id": 9445, "Python id": 9446, "finished": true, "Ev Idx": 12346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912224.752, "dur": 3.268,
    "args": {
      "Python parent id": 9446, "Python id": 9447, "finished": true, "Ev Idx": 12347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912224.994, "dur": 2.981,
    "args": {
      "Python parent id": 9447, "Python id": 9448, "finished": true, "Ev Idx": 12348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912225.125, "dur": 0.257,
    "args": {
      "Python parent id": 9448, "Python id": 9449, "finished": true, "Ev Idx": 12349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912225.254, "dur": 0.072,
    "args": {
      "Python parent id": 9449, "Python id": 9450, "finished": true, "Ev Idx": 12350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912225.573, "dur": 2.014,
    "args": {
      "Python parent id": 9448, "Python id": 9451, "finished": true, "Ev Idx": 12351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912225.791, "dur": 1.698,
    "args": {
      "Python parent id": 9451, "Python id": 9452, "finished": true, "Ev Idx": 12352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912225.998, "dur": 1.440,
    "args": {
      "Python parent id": 9452, "Python id": 9453, "finished": true, "Ev Idx": 12353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912226.137, "dur": 1.108,
    "args": {
      "Python parent id": 9453, "Python id": 9454, "finished": true, "Ev Idx": 12354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912226.305, "dur": 0.470,
    "args": {
      "Python parent id": 9454, "Python id": 9455, "finished": true, "Ev Idx": 12355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912226.405, "dur": 0.044,
    "args": {
      "Python parent id": 9455, "Python id": 9456, "finished": true, "Ev Idx": 12356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912226.594, "dur": 0.147,
    "args": {
      "Python parent id": 9455, "Python id": 9457, "finished": true, "Ev Idx": 12357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912227.762, "dur": 0.157,
    "args": {
      "Python parent id": 9448, "Python id": 9458, "finished": true, "Ev Idx": 12358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912228.159, "dur": 0.094,
    "args": {
      "Python parent id": 9446, "Python id": 9459, "finished": true, "Ev Idx": 12359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912228.500, "dur": 1.207,
    "args": {
      "Python parent id": 9446, "Python id": 9460, "finished": true, "Ev Idx": 12360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912228.650, "dur": 1.014,
    "args": {
      "Python parent id": 9460, "Python id": 9461, "finished": true, "Ev Idx": 12361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912228.791, "dur": 0.403,
    "args": {
      "Python parent id": 9461, "Python id": 9462, "finished": true, "Ev Idx": 12362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912228.963, "dur": 0.174,
    "args": {
      "Python parent id": 9462, "Python id": 9463, "finished": true, "Ev Idx": 12363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912229.047, "dur": 0.057,
    "args": {
      "Python parent id": 9463, "Python id": 9464, "finished": true, "Ev Idx": 12364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912229.342, "dur": 0.281,
    "args": {
      "Python parent id": 9461, "Python id": 9465, "finished": true, "Ev Idx": 12365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912230.067, "dur": 0.259,
    "args": {
      "Python parent id": 9430, "Python id": 9466, "finished": true, "Ev Idx": 12366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912230.953, "dur": 0.970,
    "args": {
      "Python parent id": 9430, "Python id": 9467, "finished": true, "Ev Idx": 12367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912231.454, "dur": 0.418,
    "args": {
      "Python parent id": 9467, "Python id": 9468, "finished": true, "Ev Idx": 12368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912232.417, "dur": 3.869,
    "args": {
      "Python parent id": 9429, "Python id": 9469, "finished": true, "Ev Idx": 12369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912236.939, "dur": 7.676,
    "args": {
      "Python parent id": 12, "Python id": 9470, "finished": true, "Ev Idx": 12370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912245.089, "dur": 0.104,
    "args": {
      "Python parent id": 12, "Python id": 9471, "finished": true, "Ev Idx": 12371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912245.775, "dur": 0.173,
    "args": {
      "Python parent id": 12, "Python id": 9472, "finished": true, "Ev Idx": 12372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912246.430, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 9473, "finished": true, "Ev Idx": 12373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912246.913, "dur": 9.815,
    "args": {
      "Python parent id": 12, "Python id": 9474, "finished": true, "Ev Idx": 12374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912258.061, "dur": 14.591,
    "args": {
      "Python parent id": 12, "Python id": 9475, "finished": true, "Ev Idx": 12375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912258.376, "dur": 10.884,
    "args": {
      "Python parent id": 9475, "Python id": 9476, "finished": true, "Ev Idx": 12376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912258.542, "dur": 0.491,
    "args": {
      "Python parent id": 9476, "Python id": 9477, "finished": true, "Ev Idx": 12377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912258.719, "dur": 0.234,
    "args": {
      "Python parent id": 9477, "Python id": 9478, "finished": true, "Ev Idx": 12378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912258.819, "dur": 0.098,
    "args": {
      "Python parent id": 9478, "Python id": 9479, "finished": true, "Ev Idx": 12379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912259.318, "dur": 8.136,
    "args": {
      "Python parent id": 9476, "Python id": 9480, "finished": true, "Ev Idx": 12380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912259.414, "dur": 0.072,
    "args": {
      "Python parent id": 9480, "Python id": 9481, "finished": true, "Ev Idx": 12381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912259.601, "dur": 0.060,
    "args": {
      "Python parent id": 9480, "Python id": 9482, "finished": true, "Ev Idx": 12382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912259.815, "dur": 0.059,
    "args": {
      "Python parent id": 9480, "Python id": 9483, "finished": true, "Ev Idx": 12383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912260.264, "dur": 0.071,
    "args": {
      "Python parent id": 9480, "Python id": 9484, "finished": true, "Ev Idx": 12384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912260.606, "dur": 0.087,
    "args": {
      "Python parent id": 9480, "Python id": 9485, "finished": true, "Ev Idx": 12385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912261.015, "dur": 6.369,
    "args": {
      "Python parent id": 9480, "Python id": 9486, "finished": true, "Ev Idx": 12386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912261.141, "dur": 0.065,
    "args": {
      "Python parent id": 9486, "Python id": 9487, "finished": true, "Ev Idx": 12387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912261.358, "dur": 0.060,
    "args": {
      "Python parent id": 9486, "Python id": 9488, "finished": true, "Ev Idx": 12388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912261.606, "dur": 0.064,
    "args": {
      "Python parent id": 9486, "Python id": 9489, "finished": true, "Ev Idx": 12389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912261.971, "dur": 0.070,
    "args": {
      "Python parent id": 9486, "Python id": 9490, "finished": true, "Ev Idx": 12390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912262.315, "dur": 4.992,
    "args": {
      "Python parent id": 9486, "Python id": 9491, "finished": true, "Ev Idx": 12391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912262.513, "dur": 4.733,
    "args": {
      "Python parent id": 9491, "Python id": 9492, "finished": true, "Ev Idx": 12392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912262.646, "dur": 2.893,
    "args": {
      "Python parent id": 9492, "Python id": 9493, "finished": true, "Ev Idx": 12393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912262.847, "dur": 2.649,
    "args": {
      "Python parent id": 9493, "Python id": 9494, "finished": true, "Ev Idx": 12394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912262.971, "dur": 0.231,
    "args": {
      "Python parent id": 9494, "Python id": 9495, "finished": true, "Ev Idx": 12395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912263.091, "dur": 0.062,
    "args": {
      "Python parent id": 9495, "Python id": 9496, "finished": true, "Ev Idx": 12396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912263.341, "dur": 1.838,
    "args": {
      "Python parent id": 9494, "Python id": 9497, "finished": true, "Ev Idx": 12397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912263.507, "dur": 1.599,
    "args": {
      "Python parent id": 9497, "Python id": 9498, "finished": true, "Ev Idx": 12398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912263.708, "dur": 1.344,
    "args": {
      "Python parent id": 9498, "Python id": 9499, "finished": true, "Ev Idx": 12399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912263.829, "dur": 1.049,
    "args": {
      "Python parent id": 9499, "Python id": 9500, "finished": true, "Ev Idx": 12400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912264.011, "dur": 0.449,
    "args": {
      "Python parent id": 9500, "Python id": 9501, "finished": true, "Ev Idx": 12401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912264.110, "dur": 0.045,
    "args": {
      "Python parent id": 9501, "Python id": 9502, "finished": true, "Ev Idx": 12402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912264.286, "dur": 0.129,
    "args": {
      "Python parent id": 9501, "Python id": 9503, "finished": true, "Ev Idx": 12403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912265.360, "dur": 0.103,
    "args": {
      "Python parent id": 9494, "Python id": 9504, "finished": true, "Ev Idx": 12404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912265.684, "dur": 0.106,
    "args": {
      "Python parent id": 9492, "Python id": 9505, "finished": true, "Ev Idx": 12405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.051, "dur": 1.149,
    "args": {
      "Python parent id": 9492, "Python id": 9506, "finished": true, "Ev Idx": 12406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.185, "dur": 0.969,
    "args": {
      "Python parent id": 9506, "Python id": 9507, "finished": true, "Ev Idx": 12407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.347, "dur": 0.368,
    "args": {
      "Python parent id": 9507, "Python id": 9508, "finished": true, "Ev Idx": 12408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.482, "dur": 0.173,
    "args": {
      "Python parent id": 9508, "Python id": 9509, "finished": true, "Ev Idx": 12409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.571, "dur": 0.051,
    "args": {
      "Python parent id": 9509, "Python id": 9510, "finished": true, "Ev Idx": 12410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912266.845, "dur": 0.280,
    "args": {
      "Python parent id": 9507, "Python id": 9511, "finished": true, "Ev Idx": 12411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912267.580, "dur": 0.133,
    "args": {
      "Python parent id": 9476, "Python id": 9512, "finished": true, "Ev Idx": 12412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912268.224, "dur": 0.911,
    "args": {
      "Python parent id": 9476, "Python id": 9513, "finished": true, "Ev Idx": 12413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912268.668, "dur": 0.411,
    "args": {
      "Python parent id": 9513, "Python id": 9514, "finished": true, "Ev Idx": 12414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912269.546, "dur": 3.055,
    "args": {
      "Python parent id": 9475, "Python id": 9515, "finished": true, "Ev Idx": 12415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912273.610, "dur": 109.034,
    "args": {
      "Python parent id": 12, "Python id": 9516, "finished": true, "Ev Idx": 12416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912273.964, "dur": 0.501,
    "args": {
      "Python parent id": 9516, "Python id": 9517, "finished": true, "Ev Idx": 12417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912274.303, "dur": 0.078,
    "args": {
      "Python parent id": 9517, "Python id": 9518, "finished": true, "Ev Idx": 12418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912274.753, "dur": 0.415,
    "args": {
      "Python parent id": 9516, "Python id": 9519, "finished": true, "Ev Idx": 12419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912275.445, "dur": 0.276,
    "args": {
      "Python parent id": 9516, "Python id": 9520, "finished": true, "Ev Idx": 12420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912275.933, "dur": 1.516,
    "args": {
      "Python parent id": 9516, "Python id": 9521, "finished": true, "Ev Idx": 12421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912278.012, "dur": 104.555,
    "args": {
      "Python parent id": 9516, "Python id": 9522, "finished": true, "Ev Idx": 12422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912278.844, "dur": 103.459,
    "args": {
      "Python parent id": 9522, "Python id": 9523, "finished": true, "Ev Idx": 12423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912279.201, "dur": 102.998,
    "args": {
      "Python parent id": 9523, "Python id": 9524, "finished": true, "Ev Idx": 12424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912284.431, "dur": 96.039,
    "args": {
      "Python parent id": 9524, "Python id": 9525, "finished": true, "Ev Idx": 12425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912284.844, "dur": 95.541,
    "args": {
      "Python parent id": 9525, "Python id": 9526, "finished": true, "Ev Idx": 12426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912285.430, "dur": 94.823,
    "args": {
      "Python parent id": 9526, "Python id": 9527, "finished": true, "Ev Idx": 12427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912286.061, "dur": 0.406,
    "args": {
      "Python parent id": 9527, "Python id": 9528, "finished": true, "Ev Idx": 12428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912286.316, "dur": 0.083,
    "args": {
      "Python parent id": 9528, "Python id": 9529, "finished": true, "Ev Idx": 12429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912286.772, "dur": 0.378,
    "args": {
      "Python parent id": 9527, "Python id": 9530, "finished": true, "Ev Idx": 12430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912287.049, "dur": 0.041,
    "args": {
      "Python parent id": 9530, "Python id": 9531, "finished": true, "Ev Idx": 12431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912287.486, "dur": 92.600,
    "args": {
      "Python parent id": 9527, "Python id": 9532, "finished": true, "Ev Idx": 12432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912288.112, "dur": 89.921,
    "args": {
      "Python parent id": 9532, "Python id": 9533, "finished": true, "Ev Idx": 12433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912295.151, "dur": 18.991,
    "args": {
      "Python parent id": 9533, "Python id": 9534, "finished": true, "Ev Idx": 12434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912295.246, "dur": 0.140,
    "args": {
      "Python parent id": 9534, "Python id": 9535, "finished": true, "Ev Idx": 12435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912295.696, "dur": 18.272,
    "args": {
      "Python parent id": 9534, "Python id": 9536, "finished": true, "Ev Idx": 12436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912295.891, "dur": 8.287,
    "args": {
      "Python parent id": 9536, "Python id": 9537, "finished": true, "Ev Idx": 12437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912321.871, "dur": 7.048,
    "args": {
      "Python parent id": 9533, "Python id": 9538, "finished": true, "Ev Idx": 12438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912337.295, "dur": 8.803,
    "args": {
      "Python parent id": 9533, "Python id": 9539, "finished": true, "Ev Idx": 12439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912384.603, "dur": 20.704,
    "args": {
      "Python parent id": 12, "Python id": 9540, "finished": true, "Ev Idx": 12440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912385.032, "dur": 15.050,
    "args": {
      "Python parent id": 9540, "Python id": 9541, "finished": true, "Ev Idx": 12441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912385.261, "dur": 0.707,
    "args": {
      "Python parent id": 9541, "Python id": 9542, "finished": true, "Ev Idx": 12442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912385.489, "dur": 0.390,
    "args": {
      "Python parent id": 9542, "Python id": 9543, "finished": true, "Ev Idx": 12443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912385.688, "dur": 0.149,
    "args": {
      "Python parent id": 9543, "Python id": 9544, "finished": true, "Ev Idx": 12444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912386.427, "dur": 10.622,
    "args": {
      "Python parent id": 9541, "Python id": 9545, "finished": true, "Ev Idx": 12445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912386.564, "dur": 0.167,
    "args": {
      "Python parent id": 9545, "Python id": 9546, "finished": true, "Ev Idx": 12446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912386.862, "dur": 0.078,
    "args": {
      "Python parent id": 9545, "Python id": 9547, "finished": true, "Ev Idx": 12447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912387.114, "dur": 0.070,
    "args": {
      "Python parent id": 9545, "Python id": 9548, "finished": true, "Ev Idx": 12448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912387.640, "dur": 0.098,
    "args": {
      "Python parent id": 9545, "Python id": 9549, "finished": true, "Ev Idx": 12449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912388.060, "dur": 0.082,
    "args": {
      "Python parent id": 9545, "Python id": 9550, "finished": true, "Ev Idx": 12450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912388.508, "dur": 8.477,
    "args": {
      "Python parent id": 9545, "Python id": 9551, "finished": true, "Ev Idx": 12451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912388.626, "dur": 0.065,
    "args": {
      "Python parent id": 9551, "Python id": 9552, "finished": true, "Ev Idx": 12452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912388.872, "dur": 0.068,
    "args": {
      "Python parent id": 9551, "Python id": 9553, "finished": true, "Ev Idx": 12453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912389.137, "dur": 0.076,
    "args": {
      "Python parent id": 9551, "Python id": 9554, "finished": true, "Ev Idx": 12454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912389.576, "dur": 0.044,
    "args": {
      "Python parent id": 9551, "Python id": 9555, "finished": true, "Ev Idx": 12455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912389.961, "dur": 6.927,
    "args": {
      "Python parent id": 9551, "Python id": 9556, "finished": true, "Ev Idx": 12456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912390.209, "dur": 6.621,
    "args": {
      "Python parent id": 9556, "Python id": 9557, "finished": true, "Ev Idx": 12457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912390.354, "dur": 4.383,
    "args": {
      "Python parent id": 9557, "Python id": 9558, "finished": true, "Ev Idx": 12458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912390.553, "dur": 4.116,
    "args": {
      "Python parent id": 9558, "Python id": 9559, "finished": true, "Ev Idx": 12459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912390.691, "dur": 0.361,
    "args": {
      "Python parent id": 9559, "Python id": 9560, "finished": true, "Ev Idx": 12460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912390.869, "dur": 0.094,
    "args": {
      "Python parent id": 9560, "Python id": 9561, "finished": true, "Ev Idx": 12461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912391.239, "dur": 2.996,
    "args": {
      "Python parent id": 9559, "Python id": 9562, "finished": true, "Ev Idx": 12462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912391.503, "dur": 2.561,
    "args": {
      "Python parent id": 9562, "Python id": 9563, "finished": true, "Ev Idx": 12463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912391.784, "dur": 2.195,
    "args": {
      "Python parent id": 9563, "Python id": 9564, "finished": true, "Ev Idx": 12464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912392.001, "dur": 1.636,
    "args": {
      "Python parent id": 9564, "Python id": 9565, "finished": true, "Ev Idx": 12465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912392.233, "dur": 0.651,
    "args": {
      "Python parent id": 9565, "Python id": 9566, "finished": true, "Ev Idx": 12466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912392.327, "dur": 0.054,
    "args": {
      "Python parent id": 9566, "Python id": 9567, "finished": true, "Ev Idx": 12467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912392.623, "dur": 0.212,
    "args": {
      "Python parent id": 9566, "Python id": 9568, "finished": true, "Ev Idx": 12468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912394.429, "dur": 0.180,
    "args": {
      "Python parent id": 9559, "Python id": 9569, "finished": true, "Ev Idx": 12469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912394.986, "dur": 0.119,
    "args": {
      "Python parent id": 9557, "Python id": 9570, "finished": true, "Ev Idx": 12470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912395.462, "dur": 1.310,
    "args": {
      "Python parent id": 9557, "Python id": 9571, "finished": true, "Ev Idx": 12471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912395.631, "dur": 1.083,
    "args": {
      "Python parent id": 9571, "Python id": 9572, "finished": true, "Ev Idx": 12472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912395.771, "dur": 0.435,
    "args": {
      "Python parent id": 9572, "Python id": 9573, "finished": true, "Ev Idx": 12473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912395.907, "dur": 0.222,
    "args": {
      "Python parent id": 9573, "Python id": 9574, "finished": true, "Ev Idx": 12474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912396.021, "dur": 0.075,
    "args": {
      "Python parent id": 9574, "Python id": 9575, "finished": true, "Ev Idx": 12475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912396.345, "dur": 0.320,
    "args": {
      "Python parent id": 9572, "Python id": 9576, "finished": true, "Ev Idx": 12476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912397.175, "dur": 0.342,
    "args": {
      "Python parent id": 9541, "Python id": 9577, "finished": true, "Ev Idx": 12477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912398.409, "dur": 1.500,
    "args": {
      "Python parent id": 9541, "Python id": 9578, "finished": true, "Ev Idx": 12478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912399.128, "dur": 0.709,
    "args": {
      "Python parent id": 9578, "Python id": 9579, "finished": true, "Ev Idx": 12479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912400.595, "dur": 4.618,
    "args": {
      "Python parent id": 9540, "Python id": 9580, "finished": true, "Ev Idx": 12480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912405.949, "dur": 9.352,
    "args": {
      "Python parent id": 12, "Python id": 9581, "finished": true, "Ev Idx": 12481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912415.797, "dur": 0.138,
    "args": {
      "Python parent id": 12, "Python id": 9582, "finished": true, "Ev Idx": 12482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912416.743, "dur": 0.276,
    "args": {
      "Python parent id": 12, "Python id": 9583, "finished": true, "Ev Idx": 12483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912417.558, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 9584, "finished": true, "Ev Idx": 12484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912418.032, "dur": 11.703,
    "args": {
      "Python parent id": 12, "Python id": 9585, "finished": true, "Ev Idx": 12485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912431.156, "dur": 16.735,
    "args": {
      "Python parent id": 12, "Python id": 9586, "finished": true, "Ev Idx": 12486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912431.517, "dur": 12.319,
    "args": {
      "Python parent id": 9586, "Python id": 9587, "finished": true, "Ev Idx": 12487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912431.714, "dur": 0.611,
    "args": {
      "Python parent id": 9587, "Python id": 9588, "finished": true, "Ev Idx": 12488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912431.889, "dur": 0.329,
    "args": {
      "Python parent id": 9588, "Python id": 9589, "finished": true, "Ev Idx": 12489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912432.064, "dur": 0.119,
    "args": {
      "Python parent id": 9589, "Python id": 9590, "finished": true, "Ev Idx": 12490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912432.635, "dur": 9.210,
    "args": {
      "Python parent id": 9587, "Python id": 9591, "finished": true, "Ev Idx": 12491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912432.749, "dur": 0.066,
    "args": {
      "Python parent id": 9591, "Python id": 9592, "finished": true, "Ev Idx": 12492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912432.956, "dur": 0.067,
    "args": {
      "Python parent id": 9591, "Python id": 9593, "finished": true, "Ev Idx": 12493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912433.198, "dur": 0.057,
    "args": {
      "Python parent id": 9591, "Python id": 9594, "finished": true, "Ev Idx": 12494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912433.721, "dur": 0.059,
    "args": {
      "Python parent id": 9591, "Python id": 9595, "finished": true, "Ev Idx": 12495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912434.074, "dur": 0.081,
    "args": {
      "Python parent id": 9591, "Python id": 9596, "finished": true, "Ev Idx": 12496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912434.491, "dur": 7.295,
    "args": {
      "Python parent id": 9591, "Python id": 9597, "finished": true, "Ev Idx": 12497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912434.589, "dur": 0.064,
    "args": {
      "Python parent id": 9597, "Python id": 9598, "finished": true, "Ev Idx": 12498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912434.813, "dur": 0.063,
    "args": {
      "Python parent id": 9597, "Python id": 9599, "finished": true, "Ev Idx": 12499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912435.048, "dur": 0.073,
    "args": {
      "Python parent id": 9597, "Python id": 9600, "finished": true, "Ev Idx": 12500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912435.490, "dur": 0.036,
    "args": {
      "Python parent id": 9597, "Python id": 9601, "finished": true, "Ev Idx": 12501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912435.825, "dur": 5.855,
    "args": {
      "Python parent id": 9597, "Python id": 9602, "finished": true, "Ev Idx": 12502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.080, "dur": 5.501,
    "args": {
      "Python parent id": 9602, "Python id": 9603, "finished": true, "Ev Idx": 12503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.204, "dur": 3.473,
    "args": {
      "Python parent id": 9603, "Python id": 9604, "finished": true, "Ev Idx": 12504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.409, "dur": 3.176,
    "args": {
      "Python parent id": 9604, "Python id": 9605, "finished": true, "Ev Idx": 12505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.545, "dur": 0.275,
    "args": {
      "Python parent id": 9605, "Python id": 9606, "finished": true, "Ev Idx": 12506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.672, "dur": 0.094,
    "args": {
      "Python parent id": 9606, "Python id": 9607, "finished": true, "Ev Idx": 12507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912436.986, "dur": 2.261,
    "args": {
      "Python parent id": 9605, "Python id": 9608, "finished": true, "Ev Idx": 12508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912437.166, "dur": 1.972,
    "args": {
      "Python parent id": 9608, "Python id": 9609, "finished": true, "Ev Idx": 12509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912437.375, "dur": 1.694,
    "args": {
      "Python parent id": 9609, "Python id": 9610, "finished": true, "Ev Idx": 12510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912437.542, "dur": 1.300,
    "args": {
      "Python parent id": 9610, "Python id": 9611, "finished": true, "Ev Idx": 12511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912437.747, "dur": 0.582,
    "args": {
      "Python parent id": 9611, "Python id": 9612, "finished": true, "Ev Idx": 12512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912437.841, "dur": 0.068,
    "args": {
      "Python parent id": 9612, "Python id": 9613, "finished": true, "Ev Idx": 12513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912438.097, "dur": 0.186,
    "args": {
      "Python parent id": 9612, "Python id": 9614, "finished": true, "Ev Idx": 12514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912439.432, "dur": 0.100,
    "args": {
      "Python parent id": 9605, "Python id": 9615, "finished": true, "Ev Idx": 12515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912439.877, "dur": 0.113,
    "args": {
      "Python parent id": 9603, "Python id": 9616, "finished": true, "Ev Idx": 12516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912440.271, "dur": 1.263,
    "args": {
      "Python parent id": 9603, "Python id": 9617, "finished": true, "Ev Idx": 12517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912440.449, "dur": 1.042,
    "args": {
      "Python parent id": 9617, "Python id": 9618, "finished": true, "Ev Idx": 12518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912440.606, "dur": 0.446,
    "args": {
      "Python parent id": 9618, "Python id": 9619, "finished": true, "Ev Idx": 12519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912440.790, "dur": 0.187,
    "args": {
      "Python parent id": 9619, "Python id": 9620, "finished": true, "Ev Idx": 12520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912440.870, "dur": 0.073,
    "args": {
      "Python parent id": 9620, "Python id": 9621, "finished": true, "Ev Idx": 12521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912441.186, "dur": 0.277,
    "args": {
      "Python parent id": 9618, "Python id": 9622, "finished": true, "Ev Idx": 12522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912442.018, "dur": 0.172,
    "args": {
      "Python parent id": 9587, "Python id": 9623, "finished": true, "Ev Idx": 12523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912442.761, "dur": 0.920,
    "args": {
      "Python parent id": 9587, "Python id": 9624, "finished": true, "Ev Idx": 12524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912443.164, "dur": 0.445,
    "args": {
      "Python parent id": 9624, "Python id": 9625, "finished": true, "Ev Idx": 12525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912444.170, "dur": 3.638,
    "args": {
      "Python parent id": 9586, "Python id": 9626, "finished": true, "Ev Idx": 12526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912448.852, "dur": 104.285,
    "args": {
      "Python parent id": 12, "Python id": 9627, "finished": true, "Ev Idx": 12527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912449.226, "dur": 0.538,
    "args": {
      "Python parent id": 9627, "Python id": 9628, "finished": true, "Ev Idx": 12528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912449.594, "dur": 0.123,
    "args": {
      "Python parent id": 9628, "Python id": 9629, "finished": true, "Ev Idx": 12529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912450.102, "dur": 0.539,
    "args": {
      "Python parent id": 9627, "Python id": 9630, "finished": true, "Ev Idx": 12530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912450.868, "dur": 0.275,
    "args": {
      "Python parent id": 9627, "Python id": 9631, "finished": true, "Ev Idx": 12531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912451.394, "dur": 1.939,
    "args": {
      "Python parent id": 9627, "Python id": 9632, "finished": true, "Ev Idx": 12532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912453.994, "dur": 99.071,
    "args": {
      "Python parent id": 9627, "Python id": 9633, "finished": true, "Ev Idx": 12533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912455.034, "dur": 97.769,
    "args": {
      "Python parent id": 9633, "Python id": 9634, "finished": true, "Ev Idx": 12534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912455.379, "dur": 97.363,
    "args": {
      "Python parent id": 9634, "Python id": 9635, "finished": true, "Ev Idx": 12535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912461.870, "dur": 89.399,
    "args": {
      "Python parent id": 9635, "Python id": 9636, "finished": true, "Ev Idx": 12536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912462.301, "dur": 88.865,
    "args": {
      "Python parent id": 9636, "Python id": 9637, "finished": true, "Ev Idx": 12537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912462.917, "dur": 88.094,
    "args": {
      "Python parent id": 9637, "Python id": 9638, "finished": true, "Ev Idx": 12538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912463.512, "dur": 0.466,
    "args": {
      "Python parent id": 9638, "Python id": 9639, "finished": true, "Ev Idx": 12539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912463.826, "dur": 0.079,
    "args": {
      "Python parent id": 9639, "Python id": 9640, "finished": true, "Ev Idx": 12540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912464.319, "dur": 0.395,
    "args": {
      "Python parent id": 9638, "Python id": 9641, "finished": true, "Ev Idx": 12541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912464.594, "dur": 0.071,
    "args": {
      "Python parent id": 9641, "Python id": 9642, "finished": true, "Ev Idx": 12542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912465.085, "dur": 85.726,
    "args": {
      "Python parent id": 9638, "Python id": 9643, "finished": true, "Ev Idx": 12543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912465.888, "dur": 83.435,
    "args": {
      "Python parent id": 9643, "Python id": 9644, "finished": true, "Ev Idx": 12544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912474.896, "dur": 20.012,
    "args": {
      "Python parent id": 9644, "Python id": 9645, "finished": true, "Ev Idx": 12545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912475.053, "dur": 0.127,
    "args": {
      "Python parent id": 9645, "Python id": 9646, "finished": true, "Ev Idx": 12546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912475.526, "dur": 19.209,
    "args": {
      "Python parent id": 9645, "Python id": 9647, "finished": true, "Ev Idx": 12547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912475.668, "dur": 8.532,
    "args": {
      "Python parent id": 9647, "Python id": 9648, "finished": true, "Ev Idx": 12548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912504.597, "dur": 7.219,
    "args": {
      "Python parent id": 9644, "Python id": 9649, "finished": true, "Ev Idx": 12549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912523.142, "dur": 9.553,
    "args": {
      "Python parent id": 9644, "Python id": 9650, "finished": true, "Ev Idx": 12550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912554.425, "dur": 18.260,
    "args": {
      "Python parent id": 12, "Python id": 9651, "finished": true, "Ev Idx": 12551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912554.815, "dur": 14.028,
    "args": {
      "Python parent id": 9651, "Python id": 9652, "finished": true, "Ev Idx": 12552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912555.001, "dur": 0.610,
    "args": {
      "Python parent id": 9652, "Python id": 9653, "finished": true, "Ev Idx": 12553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912555.196, "dur": 0.325,
    "args": {
      "Python parent id": 9653, "Python id": 9654, "finished": true, "Ev Idx": 12554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912555.365, "dur": 0.121,
    "args": {
      "Python parent id": 9654, "Python id": 9655, "finished": true, "Ev Idx": 12555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912555.946, "dur": 10.352,
    "args": {
      "Python parent id": 9652, "Python id": 9656, "finished": true, "Ev Idx": 12556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912556.047, "dur": 0.110,
    "args": {
      "Python parent id": 9656, "Python id": 9657, "finished": true, "Ev Idx": 12557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912556.300, "dur": 0.075,
    "args": {
      "Python parent id": 9656, "Python id": 9658, "finished": true, "Ev Idx": 12558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912556.554, "dur": 0.060,
    "args": {
      "Python parent id": 9656, "Python id": 9659, "finished": true, "Ev Idx": 12559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912557.119, "dur": 0.069,
    "args": {
      "Python parent id": 9656, "Python id": 9660, "finished": true, "Ev Idx": 12560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912557.544, "dur": 0.075,
    "args": {
      "Python parent id": 9656, "Python id": 9661, "finished": true, "Ev Idx": 12561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912558.022, "dur": 8.202,
    "args": {
      "Python parent id": 9656, "Python id": 9662, "finished": true, "Ev Idx": 12562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912558.134, "dur": 0.061,
    "args": {
      "Python parent id": 9662, "Python id": 9663, "finished": true, "Ev Idx": 12563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912558.396, "dur": 0.064,
    "args": {
      "Python parent id": 9662, "Python id": 9664, "finished": true, "Ev Idx": 12564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912558.670, "dur": 0.072,
    "args": {
      "Python parent id": 9662, "Python id": 9665, "finished": true, "Ev Idx": 12565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912559.146, "dur": 0.039,
    "args": {
      "Python parent id": 9662, "Python id": 9666, "finished": true, "Ev Idx": 12566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912559.539, "dur": 6.602,
    "args": {
      "Python parent id": 9662, "Python id": 9667, "finished": true, "Ev Idx": 12567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912559.820, "dur": 6.216,
    "args": {
      "Python parent id": 9667, "Python id": 9668, "finished": true, "Ev Idx": 12568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912559.957, "dur": 4.033,
    "args": {
      "Python parent id": 9668, "Python id": 9669, "finished": true, "Ev Idx": 12569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912560.155, "dur": 3.770,
    "args": {
      "Python parent id": 9669, "Python id": 9670, "finished": true, "Ev Idx": 12570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912560.293, "dur": 0.278,
    "args": {
      "Python parent id": 9670, "Python id": 9671, "finished": true, "Ev Idx": 12571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912560.426, "dur": 0.068,
    "args": {
      "Python parent id": 9671, "Python id": 9672, "finished": true, "Ev Idx": 12572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912560.769, "dur": 2.738,
    "args": {
      "Python parent id": 9670, "Python id": 9673, "finished": true, "Ev Idx": 12573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912561.037, "dur": 2.354,
    "args": {
      "Python parent id": 9673, "Python id": 9674, "finished": true, "Ev Idx": 12574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912561.281, "dur": 2.021,
    "args": {
      "Python parent id": 9674, "Python id": 9675, "finished": true, "Ev Idx": 12575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912561.513, "dur": 1.504,
    "args": {
      "Python parent id": 9675, "Python id": 9676, "finished": true, "Ev Idx": 12576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912561.768, "dur": 0.551,
    "args": {
      "Python parent id": 9676, "Python id": 9677, "finished": true, "Ev Idx": 12577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912561.881, "dur": 0.052,
    "args": {
      "Python parent id": 9677, "Python id": 9678, "finished": true, "Ev Idx": 12578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912562.111, "dur": 0.167,
    "args": {
      "Python parent id": 9677, "Python id": 9679, "finished": true, "Ev Idx": 12579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912563.701, "dur": 0.181,
    "args": {
      "Python parent id": 9670, "Python id": 9680, "finished": true, "Ev Idx": 12580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912564.220, "dur": 0.096,
    "args": {
      "Python parent id": 9668, "Python id": 9681, "finished": true, "Ev Idx": 12581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912564.621, "dur": 1.368,
    "args": {
      "Python parent id": 9668, "Python id": 9682, "finished": true, "Ev Idx": 12582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912564.849, "dur": 1.093,
    "args": {
      "Python parent id": 9682, "Python id": 9683, "finished": true, "Ev Idx": 12583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912565.009, "dur": 0.460,
    "args": {
      "Python parent id": 9683, "Python id": 9684, "finished": true, "Ev Idx": 12584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912565.174, "dur": 0.220,
    "args": {
      "Python parent id": 9684, "Python id": 9685, "finished": true, "Ev Idx": 12585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912565.285, "dur": 0.079,
    "args": {
      "Python parent id": 9685, "Python id": 9686, "finished": true, "Ev Idx": 12586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912565.613, "dur": 0.282,
    "args": {
      "Python parent id": 9683, "Python id": 9687, "finished": true, "Ev Idx": 12587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912566.485, "dur": 0.263,
    "args": {
      "Python parent id": 9652, "Python id": 9688, "finished": true, "Ev Idx": 12588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912567.445, "dur": 1.220,
    "args": {
      "Python parent id": 9652, "Python id": 9689, "finished": true, "Ev Idx": 12589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912568.037, "dur": 0.572,
    "args": {
      "Python parent id": 9689, "Python id": 9690, "finished": true, "Ev Idx": 12590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912569.192, "dur": 3.387,
    "args": {
      "Python parent id": 9651, "Python id": 9691, "finished": true, "Ev Idx": 12591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912573.317, "dur": 9.085,
    "args": {
      "Python parent id": 12, "Python id": 9692, "finished": true, "Ev Idx": 12592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912582.869, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 9693, "finished": true, "Ev Idx": 12593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912583.613, "dur": 0.196,
    "args": {
      "Python parent id": 12, "Python id": 9694, "finished": true, "Ev Idx": 12594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912584.301, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 9695, "finished": true, "Ev Idx": 12595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912584.773, "dur": 11.038,
    "args": {
      "Python parent id": 12, "Python id": 9696, "finished": true, "Ev Idx": 12596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912597.231, "dur": 16.091,
    "args": {
      "Python parent id": 12, "Python id": 9697, "finished": true, "Ev Idx": 12597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912597.635, "dur": 12.141,
    "args": {
      "Python parent id": 9697, "Python id": 9698, "finished": true, "Ev Idx": 12598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912597.804, "dur": 0.495,
    "args": {
      "Python parent id": 9698, "Python id": 9699, "finished": true, "Ev Idx": 12599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912597.958, "dur": 0.238,
    "args": {
      "Python parent id": 9699, "Python id": 9700, "finished": true, "Ev Idx": 12600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912598.068, "dur": 0.094,
    "args": {
      "Python parent id": 9700, "Python id": 9701, "finished": true, "Ev Idx": 12601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912598.570, "dur": 9.176,
    "args": {
      "Python parent id": 9698, "Python id": 9702, "finished": true, "Ev Idx": 12602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912598.661, "dur": 0.114,
    "args": {
      "Python parent id": 9702, "Python id": 9703, "finished": true, "Ev Idx": 12603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912598.885, "dur": 0.069,
    "args": {
      "Python parent id": 9702, "Python id": 9704, "finished": true, "Ev Idx": 12604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912599.106, "dur": 0.064,
    "args": {
      "Python parent id": 9702, "Python id": 9705, "finished": true, "Ev Idx": 12605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912599.587, "dur": 0.077,
    "args": {
      "Python parent id": 9702, "Python id": 9706, "finished": true, "Ev Idx": 12606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912599.979, "dur": 0.083,
    "args": {
      "Python parent id": 9702, "Python id": 9707, "finished": true, "Ev Idx": 12607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912600.481, "dur": 7.177,
    "args": {
      "Python parent id": 9702, "Python id": 9708, "finished": true, "Ev Idx": 12608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912600.563, "dur": 0.073,
    "args": {
      "Python parent id": 9708, "Python id": 9709, "finished": true, "Ev Idx": 12609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912600.787, "dur": 0.066,
    "args": {
      "Python parent id": 9708, "Python id": 9710, "finished": true, "Ev Idx": 12610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912601.053, "dur": 0.075,
    "args": {
      "Python parent id": 9708, "Python id": 9711, "finished": true, "Ev Idx": 12611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912601.468, "dur": 0.045,
    "args": {
      "Python parent id": 9708, "Python id": 9712, "finished": true, "Ev Idx": 12612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912601.840, "dur": 5.695,
    "args": {
      "Python parent id": 9708, "Python id": 9713, "finished": true, "Ev Idx": 12613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912602.082, "dur": 5.366,
    "args": {
      "Python parent id": 9713, "Python id": 9714, "finished": true, "Ev Idx": 12614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912602.220, "dur": 3.451,
    "args": {
      "Python parent id": 9714, "Python id": 9715, "finished": true, "Ev Idx": 12615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912602.441, "dur": 3.150,
    "args": {
      "Python parent id": 9715, "Python id": 9716, "finished": true, "Ev Idx": 12616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912602.576, "dur": 0.267,
    "args": {
      "Python parent id": 9716, "Python id": 9717, "finished": true, "Ev Idx": 12617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912602.695, "dur": 0.073,
    "args": {
      "Python parent id": 9717, "Python id": 9718, "finished": true, "Ev Idx": 12618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.029, "dur": 2.193,
    "args": {
      "Python parent id": 9716, "Python id": 9719, "finished": true, "Ev Idx": 12619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.238, "dur": 1.896,
    "args": {
      "Python parent id": 9719, "Python id": 9720, "finished": true, "Ev Idx": 12620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.456, "dur": 1.628,
    "args": {
      "Python parent id": 9720, "Python id": 9721, "finished": true, "Ev Idx": 12621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.599, "dur": 1.255,
    "args": {
      "Python parent id": 9721, "Python id": 9722, "finished": true, "Ev Idx": 12622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.783, "dur": 0.539,
    "args": {
      "Python parent id": 9722, "Python id": 9723, "finished": true, "Ev Idx": 12623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912603.870, "dur": 0.071,
    "args": {
      "Python parent id": 9723, "Python id": 9724, "finished": true, "Ev Idx": 12624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912604.133, "dur": 0.132,
    "args": {
      "Python parent id": 9723, "Python id": 9725, "finished": true, "Ev Idx": 12625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912605.434, "dur": 0.105,
    "args": {
      "Python parent id": 9716, "Python id": 9726, "finished": true, "Ev Idx": 12626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912605.839, "dur": 0.081,
    "args": {
      "Python parent id": 9714, "Python id": 9727, "finished": true, "Ev Idx": 12627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912606.186, "dur": 1.218,
    "args": {
      "Python parent id": 9714, "Python id": 9728, "finished": true, "Ev Idx": 12628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912606.327, "dur": 1.035,
    "args": {
      "Python parent id": 9728, "Python id": 9729, "finished": true, "Ev Idx": 12629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912606.495, "dur": 0.395,
    "args": {
      "Python parent id": 9729, "Python id": 9730, "finished": true, "Ev Idx": 12630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912606.649, "dur": 0.182,
    "args": {
      "Python parent id": 9730, "Python id": 9731, "finished": true, "Ev Idx": 12631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912606.747, "dur": 0.051,
    "args": {
      "Python parent id": 9731, "Python id": 9732, "finished": true, "Ev Idx": 12632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912607.027, "dur": 0.291,
    "args": {
      "Python parent id": 9729, "Python id": 9733, "finished": true, "Ev Idx": 12633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912607.874, "dur": 0.201,
    "args": {
      "Python parent id": 9698, "Python id": 9734, "finished": true, "Ev Idx": 12634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912608.710, "dur": 0.943,
    "args": {
      "Python parent id": 9698, "Python id": 9735, "finished": true, "Ev Idx": 12635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912609.145, "dur": 0.450,
    "args": {
      "Python parent id": 9735, "Python id": 9736, "finished": true, "Ev Idx": 12636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912610.045, "dur": 3.231,
    "args": {
      "Python parent id": 9697, "Python id": 9737, "finished": true, "Ev Idx": 12637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912614.254, "dur": 97.975,
    "args": {
      "Python parent id": 12, "Python id": 9738, "finished": true, "Ev Idx": 12638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912614.732, "dur": 0.444,
    "args": {
      "Python parent id": 9738, "Python id": 9739, "finished": true, "Ev Idx": 12639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912615.012, "dur": 0.115,
    "args": {
      "Python parent id": 9739, "Python id": 9740, "finished": true, "Ev Idx": 12640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912615.822, "dur": 0.442,
    "args": {
      "Python parent id": 9738, "Python id": 9741, "finished": true, "Ev Idx": 12641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912616.653, "dur": 0.265,
    "args": {
      "Python parent id": 9738, "Python id": 9742, "finished": true, "Ev Idx": 12642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912617.323, "dur": 1.680,
    "args": {
      "Python parent id": 9738, "Python id": 9743, "finished": true, "Ev Idx": 12643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912619.657, "dur": 92.501,
    "args": {
      "Python parent id": 9738, "Python id": 9744, "finished": true, "Ev Idx": 12644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912620.794, "dur": 91.128,
    "args": {
      "Python parent id": 9744, "Python id": 9745, "finished": true, "Ev Idx": 12645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912621.182, "dur": 90.685,
    "args": {
      "Python parent id": 9745, "Python id": 9746, "finished": true, "Ev Idx": 12646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912627.337, "dur": 83.048,
    "args": {
      "Python parent id": 9746, "Python id": 9747, "finished": true, "Ev Idx": 12647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912627.687, "dur": 82.646,
    "args": {
      "Python parent id": 9747, "Python id": 9748, "finished": true, "Ev Idx": 12648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912628.283, "dur": 81.881,
    "args": {
      "Python parent id": 9748, "Python id": 9749, "finished": true, "Ev Idx": 12649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912628.979, "dur": 0.472,
    "args": {
      "Python parent id": 9749, "Python id": 9750, "finished": true, "Ev Idx": 12650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912629.298, "dur": 0.047,
    "args": {
      "Python parent id": 9750, "Python id": 9751, "finished": true, "Ev Idx": 12651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912629.917, "dur": 0.341,
    "args": {
      "Python parent id": 9749, "Python id": 9752, "finished": true, "Ev Idx": 12652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912630.163, "dur": 0.042,
    "args": {
      "Python parent id": 9752, "Python id": 9753, "finished": true, "Ev Idx": 12653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912630.608, "dur": 79.331,
    "args": {
      "Python parent id": 9749, "Python id": 9754, "finished": true, "Ev Idx": 12654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912631.749, "dur": 76.697,
    "args": {
      "Python parent id": 9754, "Python id": 9755, "finished": true, "Ev Idx": 12655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912639.475, "dur": 19.329,
    "args": {
      "Python parent id": 9755, "Python id": 9756, "finished": true, "Ev Idx": 12656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912639.719, "dur": 0.158,
    "args": {
      "Python parent id": 9756, "Python id": 9757, "finished": true, "Ev Idx": 12657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912640.262, "dur": 18.343,
    "args": {
      "Python parent id": 9756, "Python id": 9758, "finished": true, "Ev Idx": 12658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912640.427, "dur": 7.565,
    "args": {
      "Python parent id": 9758, "Python id": 9759, "finished": true, "Ev Idx": 12659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912667.313, "dur": 7.235,
    "args": {
      "Python parent id": 9755, "Python id": 9760, "finished": true, "Ev Idx": 12660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912683.707, "dur": 8.996,
    "args": {
      "Python parent id": 9755, "Python id": 9761, "finished": true, "Ev Idx": 12661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912713.499, "dur": 18.584,
    "args": {
      "Python parent id": 12, "Python id": 9762, "finished": true, "Ev Idx": 12662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912713.955, "dur": 13.909,
    "args": {
      "Python parent id": 9762, "Python id": 9763, "finished": true, "Ev Idx": 12663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912714.139, "dur": 0.656,
    "args": {
      "Python parent id": 9763, "Python id": 9764, "finished": true, "Ev Idx": 12664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912714.349, "dur": 0.315,
    "args": {
      "Python parent id": 9764, "Python id": 9765, "finished": true, "Ev Idx": 12665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912714.509, "dur": 0.120,
    "args": {
      "Python parent id": 9765, "Python id": 9766, "finished": true, "Ev Idx": 12666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912715.221, "dur": 10.178,
    "args": {
      "Python parent id": 9763, "Python id": 9767, "finished": true, "Ev Idx": 12667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912715.333, "dur": 0.066,
    "args": {
      "Python parent id": 9767, "Python id": 9768, "finished": true, "Ev Idx": 12668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912715.541, "dur": 0.079,
    "args": {
      "Python parent id": 9767, "Python id": 9769, "finished": true, "Ev Idx": 12669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912715.817, "dur": 0.059,
    "args": {
      "Python parent id": 9767, "Python id": 9770, "finished": true, "Ev Idx": 12670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912716.303, "dur": 0.038,
    "args": {
      "Python parent id": 9767, "Python id": 9771, "finished": true, "Ev Idx": 12671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912716.625, "dur": 0.090,
    "args": {
      "Python parent id": 9767, "Python id": 9772, "finished": true, "Ev Idx": 12672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912717.111, "dur": 8.222,
    "args": {
      "Python parent id": 9767, "Python id": 9773, "finished": true, "Ev Idx": 12673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912717.216, "dur": 0.064,
    "args": {
      "Python parent id": 9773, "Python id": 9774, "finished": true, "Ev Idx": 12674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912717.437, "dur": 0.058,
    "args": {
      "Python parent id": 9773, "Python id": 9775, "finished": true, "Ev Idx": 12675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912717.717, "dur": 0.063,
    "args": {
      "Python parent id": 9773, "Python id": 9776, "finished": true, "Ev Idx": 12676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912718.138, "dur": 0.055,
    "args": {
      "Python parent id": 9773, "Python id": 9777, "finished": true, "Ev Idx": 12677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912718.520, "dur": 6.713,
    "args": {
      "Python parent id": 9773, "Python id": 9778, "finished": true, "Ev Idx": 12678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912718.739, "dur": 6.392,
    "args": {
      "Python parent id": 9778, "Python id": 9779, "finished": true, "Ev Idx": 12679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912718.923, "dur": 4.031,
    "args": {
      "Python parent id": 9779, "Python id": 9780, "finished": true, "Ev Idx": 12680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912719.119, "dur": 3.764,
    "args": {
      "Python parent id": 9780, "Python id": 9781, "finished": true, "Ev Idx": 12681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912719.291, "dur": 0.311,
    "args": {
      "Python parent id": 9781, "Python id": 9782, "finished": true, "Ev Idx": 12682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912719.445, "dur": 0.094,
    "args": {
      "Python parent id": 9782, "Python id": 9783, "finished": true, "Ev Idx": 12683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912719.826, "dur": 2.620,
    "args": {
      "Python parent id": 9781, "Python id": 9784, "finished": true, "Ev Idx": 12684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912720.108, "dur": 2.207,
    "args": {
      "Python parent id": 9784, "Python id": 9785, "finished": true, "Ev Idx": 12685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912720.323, "dur": 1.924,
    "args": {
      "Python parent id": 9785, "Python id": 9786, "finished": true, "Ev Idx": 12686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912720.529, "dur": 1.457,
    "args": {
      "Python parent id": 9786, "Python id": 9787, "finished": true, "Ev Idx": 12687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912720.808, "dur": 0.567,
    "args": {
      "Python parent id": 9787, "Python id": 9788, "finished": true, "Ev Idx": 12688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912720.932, "dur": 0.061,
    "args": {
      "Python parent id": 9788, "Python id": 9789, "finished": true, "Ev Idx": 12689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912721.166, "dur": 0.151,
    "args": {
      "Python parent id": 9788, "Python id": 9790, "finished": true, "Ev Idx": 12690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912722.647, "dur": 0.166,
    "args": {
      "Python parent id": 9781, "Python id": 9791, "finished": true, "Ev Idx": 12691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912723.220, "dur": 0.097,
    "args": {
      "Python parent id": 9779, "Python id": 9792, "finished": true, "Ev Idx": 12692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912723.609, "dur": 1.477,
    "args": {
      "Python parent id": 9779, "Python id": 9793, "finished": true, "Ev Idx": 12693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912723.804, "dur": 1.237,
    "args": {
      "Python parent id": 9793, "Python id": 9794, "finished": true, "Ev Idx": 12694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912723.968, "dur": 0.532,
    "args": {
      "Python parent id": 9794, "Python id": 9795, "finished": true, "Ev Idx": 12695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912724.152, "dur": 0.261,
    "args": {
      "Python parent id": 9795, "Python id": 9796, "finished": true, "Ev Idx": 12696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912724.263, "dur": 0.087,
    "args": {
      "Python parent id": 9796, "Python id": 9797, "finished": true, "Ev Idx": 12697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912724.697, "dur": 0.301,
    "args": {
      "Python parent id": 9794, "Python id": 9798, "finished": true, "Ev Idx": 12698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912725.554, "dur": 0.257,
    "args": {
      "Python parent id": 9763, "Python id": 9799, "finished": true, "Ev Idx": 12699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912726.460, "dur": 1.254,
    "args": {
      "Python parent id": 9763, "Python id": 9800, "finished": true, "Ev Idx": 12700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912727.046, "dur": 0.616,
    "args": {
      "Python parent id": 9800, "Python id": 9801, "finished": true, "Ev Idx": 12701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912728.259, "dur": 3.709,
    "args": {
      "Python parent id": 9762, "Python id": 9802, "finished": true, "Ev Idx": 12702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912732.699, "dur": 9.069,
    "args": {
      "Python parent id": 12, "Python id": 9803, "finished": true, "Ev Idx": 12703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912742.221, "dur": 0.083,
    "args": {
      "Python parent id": 12, "Python id": 9804, "finished": true, "Ev Idx": 12704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912742.956, "dur": 0.225,
    "args": {
      "Python parent id": 12, "Python id": 9805, "finished": true, "Ev Idx": 12705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912743.662, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 9806, "finished": true, "Ev Idx": 12706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912744.152, "dur": 10.176,
    "args": {
      "Python parent id": 12, "Python id": 9807, "finished": true, "Ev Idx": 12707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912755.766, "dur": 16.084,
    "args": {
      "Python parent id": 12, "Python id": 9808, "finished": true, "Ev Idx": 12708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912756.104, "dur": 12.359,
    "args": {
      "Python parent id": 9808, "Python id": 9809, "finished": true, "Ev Idx": 12709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912756.293, "dur": 0.514,
    "args": {
      "Python parent id": 9809, "Python id": 9810, "finished": true, "Ev Idx": 12710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912756.493, "dur": 0.246,
    "args": {
      "Python parent id": 9810, "Python id": 9811, "finished": true, "Ev Idx": 12711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912756.581, "dur": 0.098,
    "args": {
      "Python parent id": 9811, "Python id": 9812, "finished": true, "Ev Idx": 12712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912757.110, "dur": 9.363,
    "args": {
      "Python parent id": 9809, "Python id": 9813, "finished": true, "Ev Idx": 12713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912757.194, "dur": 0.107,
    "args": {
      "Python parent id": 9813, "Python id": 9814, "finished": true, "Ev Idx": 12714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912757.460, "dur": 0.065,
    "args": {
      "Python parent id": 9813, "Python id": 9815, "finished": true, "Ev Idx": 12715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912757.698, "dur": 0.063,
    "args": {
      "Python parent id": 9813, "Python id": 9816, "finished": true, "Ev Idx": 12716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912758.188, "dur": 0.049,
    "args": {
      "Python parent id": 9813, "Python id": 9817, "finished": true, "Ev Idx": 12717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912758.590, "dur": 0.082,
    "args": {
      "Python parent id": 9813, "Python id": 9818, "finished": true, "Ev Idx": 12718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912759.046, "dur": 7.364,
    "args": {
      "Python parent id": 9813, "Python id": 9819, "finished": true, "Ev Idx": 12719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912759.138, "dur": 0.065,
    "args": {
      "Python parent id": 9819, "Python id": 9820, "finished": true, "Ev Idx": 12720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912759.355, "dur": 0.059,
    "args": {
      "Python parent id": 9819, "Python id": 9821, "finished": true, "Ev Idx": 12721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912759.591, "dur": 0.072,
    "args": {
      "Python parent id": 9819, "Python id": 9822, "finished": true, "Ev Idx": 12722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912760.050, "dur": 0.030,
    "args": {
      "Python parent id": 9819, "Python id": 9823, "finished": true, "Ev Idx": 12723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912760.386, "dur": 5.922,
    "args": {
      "Python parent id": 9819, "Python id": 9824, "finished": true, "Ev Idx": 12724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912760.637, "dur": 5.608,
    "args": {
      "Python parent id": 9824, "Python id": 9825, "finished": true, "Ev Idx": 12725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912760.778, "dur": 3.715,
    "args": {
      "Python parent id": 9825, "Python id": 9826, "finished": true, "Ev Idx": 12726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912760.995, "dur": 3.426,
    "args": {
      "Python parent id": 9826, "Python id": 9827, "finished": true, "Ev Idx": 12727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912761.188, "dur": 0.229,
    "args": {
      "Python parent id": 9827, "Python id": 9828, "finished": true, "Ev Idx": 12728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912761.301, "dur": 0.071,
    "args": {
      "Python parent id": 9828, "Python id": 9829, "finished": true, "Ev Idx": 12729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912761.618, "dur": 2.463,
    "args": {
      "Python parent id": 9827, "Python id": 9830, "finished": true, "Ev Idx": 12730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912761.809, "dur": 2.162,
    "args": {
      "Python parent id": 9830, "Python id": 9831, "finished": true, "Ev Idx": 12731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912762.052, "dur": 1.852,
    "args": {
      "Python parent id": 9831, "Python id": 9832, "finished": true, "Ev Idx": 12732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912762.190, "dur": 1.543,
    "args": {
      "Python parent id": 9832, "Python id": 9833, "finished": true, "Ev Idx": 12733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912762.342, "dur": 0.905,
    "args": {
      "Python parent id": 9833, "Python id": 9834, "finished": true, "Ev Idx": 12734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912762.439, "dur": 0.065,
    "args": {
      "Python parent id": 9834, "Python id": 9835, "finished": true, "Ev Idx": 12735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912763.020, "dur": 0.148,
    "args": {
      "Python parent id": 9834, "Python id": 9836, "finished": true, "Ev Idx": 12736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912764.278, "dur": 0.082,
    "args": {
      "Python parent id": 9827, "Python id": 9837, "finished": true, "Ev Idx": 12737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912764.649, "dur": 0.113,
    "args": {
      "Python parent id": 9825, "Python id": 9838, "finished": true, "Ev Idx": 12738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.046, "dur": 1.151,
    "args": {
      "Python parent id": 9825, "Python id": 9839, "finished": true, "Ev Idx": 12739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.180, "dur": 0.970,
    "args": {
      "Python parent id": 9839, "Python id": 9840, "finished": true, "Ev Idx": 12740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.365, "dur": 0.382,
    "args": {
      "Python parent id": 9840, "Python id": 9841, "finished": true, "Ev Idx": 12741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.524, "dur": 0.162,
    "args": {
      "Python parent id": 9841, "Python id": 9842, "finished": true, "Ev Idx": 12742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.601, "dur": 0.056,
    "args": {
      "Python parent id": 9842, "Python id": 9843, "finished": true, "Ev Idx": 12743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912765.873, "dur": 0.245,
    "args": {
      "Python parent id": 9840, "Python id": 9844, "finished": true, "Ev Idx": 12744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912766.619, "dur": 0.149,
    "args": {
      "Python parent id": 9809, "Python id": 9845, "finished": true, "Ev Idx": 12745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912767.345, "dur": 0.993,
    "args": {
      "Python parent id": 9809, "Python id": 9846, "finished": true, "Ev Idx": 12746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912767.826, "dur": 0.460,
    "args": {
      "Python parent id": 9846, "Python id": 9847, "finished": true, "Ev Idx": 12747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912768.751, "dur": 3.014,
    "args": {
      "Python parent id": 9808, "Python id": 9848, "finished": true, "Ev Idx": 12748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912772.822, "dur": 92.728,
    "args": {
      "Python parent id": 12, "Python id": 9849, "finished": true, "Ev Idx": 12749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912773.188, "dur": 0.488,
    "args": {
      "Python parent id": 9849, "Python id": 9850, "finished": true, "Ev Idx": 12750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912773.520, "dur": 0.110,
    "args": {
      "Python parent id": 9850, "Python id": 9851, "finished": true, "Ev Idx": 12751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912773.984, "dur": 0.420,
    "args": {
      "Python parent id": 9849, "Python id": 9852, "finished": true, "Ev Idx": 12752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912774.616, "dur": 0.225,
    "args": {
      "Python parent id": 9849, "Python id": 9853, "finished": true, "Ev Idx": 12753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912775.108, "dur": 1.656,
    "args": {
      "Python parent id": 9849, "Python id": 9854, "finished": true, "Ev Idx": 12754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912777.448, "dur": 88.030,
    "args": {
      "Python parent id": 9849, "Python id": 9855, "finished": true, "Ev Idx": 12755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912778.498, "dur": 86.766,
    "args": {
      "Python parent id": 9855, "Python id": 9856, "finished": true, "Ev Idx": 12756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912778.907, "dur": 86.288,
    "args": {
      "Python parent id": 9856, "Python id": 9857, "finished": true, "Ev Idx": 12757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912784.368, "dur": 79.464,
    "args": {
      "Python parent id": 9857, "Python id": 9858, "finished": true, "Ev Idx": 12758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912784.729, "dur": 79.025,
    "args": {
      "Python parent id": 9858, "Python id": 9859, "finished": true, "Ev Idx": 12759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912785.363, "dur": 78.235,
    "args": {
      "Python parent id": 9859, "Python id": 9860, "finished": true, "Ev Idx": 12760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912785.924, "dur": 0.354,
    "args": {
      "Python parent id": 9860, "Python id": 9861, "finished": true, "Ev Idx": 12761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912786.162, "dur": 0.064,
    "args": {
      "Python parent id": 9861, "Python id": 9862, "finished": true, "Ev Idx": 12762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912786.589, "dur": 0.327,
    "args": {
      "Python parent id": 9860, "Python id": 9863, "finished": true, "Ev Idx": 12763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912786.817, "dur": 0.041,
    "args": {
      "Python parent id": 9863, "Python id": 9864, "finished": true, "Ev Idx": 12764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912787.264, "dur": 76.132,
    "args": {
      "Python parent id": 9860, "Python id": 9865, "finished": true, "Ev Idx": 12765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912788.046, "dur": 73.852,
    "args": {
      "Python parent id": 9865, "Python id": 9866, "finished": true, "Ev Idx": 12766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912795.430, "dur": 18.736,
    "args": {
      "Python parent id": 9866, "Python id": 9867, "finished": true, "Ev Idx": 12767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912795.563, "dur": 0.135,
    "args": {
      "Python parent id": 9867, "Python id": 9868, "finished": true, "Ev Idx": 12768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912796.019, "dur": 18.029,
    "args": {
      "Python parent id": 9867, "Python id": 9869, "finished": true, "Ev Idx": 12769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912796.157, "dur": 8.475,
    "args": {
      "Python parent id": 9869, "Python id": 9870, "finished": true, "Ev Idx": 12770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912822.202, "dur": 7.167,
    "args": {
      "Python parent id": 9866, "Python id": 9871, "finished": true, "Ev Idx": 12771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912837.596, "dur": 8.641,
    "args": {
      "Python parent id": 9866, "Python id": 9872, "finished": true, "Ev Idx": 12772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912866.727, "dur": 18.236,
    "args": {
      "Python parent id": 12, "Python id": 9873, "finished": true, "Ev Idx": 12773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912867.138, "dur": 13.969,
    "args": {
      "Python parent id": 9873, "Python id": 9874, "finished": true, "Ev Idx": 12774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912867.345, "dur": 0.561,
    "args": {
      "Python parent id": 9874, "Python id": 9875, "finished": true, "Ev Idx": 12775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912867.526, "dur": 0.272,
    "args": {
      "Python parent id": 9875, "Python id": 9876, "finished": true, "Ev Idx": 12776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912867.646, "dur": 0.116,
    "args": {
      "Python parent id": 9876, "Python id": 9877, "finished": true, "Ev Idx": 12777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912868.285, "dur": 10.387,
    "args": {
      "Python parent id": 9874, "Python id": 9878, "finished": true, "Ev Idx": 12778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912868.450, "dur": 0.066,
    "args": {
      "Python parent id": 9878, "Python id": 9879, "finished": true, "Ev Idx": 12779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912868.665, "dur": 0.082,
    "args": {
      "Python parent id": 9878, "Python id": 9880, "finished": true, "Ev Idx": 12780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912868.901, "dur": 0.058,
    "args": {
      "Python parent id": 9878, "Python id": 9881, "finished": true, "Ev Idx": 12781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912869.412, "dur": 0.067,
    "args": {
      "Python parent id": 9878, "Python id": 9882, "finished": true, "Ev Idx": 12782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912869.837, "dur": 0.118,
    "args": {
      "Python parent id": 9878, "Python id": 9883, "finished": true, "Ev Idx": 12783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912870.330, "dur": 8.268,
    "args": {
      "Python parent id": 9878, "Python id": 9884, "finished": true, "Ev Idx": 12784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912870.453, "dur": 0.065,
    "args": {
      "Python parent id": 9884, "Python id": 9885, "finished": true, "Ev Idx": 12785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912870.703, "dur": 0.065,
    "args": {
      "Python parent id": 9884, "Python id": 9886, "finished": true, "Ev Idx": 12786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912870.990, "dur": 0.070,
    "args": {
      "Python parent id": 9884, "Python id": 9887, "finished": true, "Ev Idx": 12787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912871.468, "dur": 0.051,
    "args": {
      "Python parent id": 9884, "Python id": 9888, "finished": true, "Ev Idx": 12788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912871.830, "dur": 6.669,
    "args": {
      "Python parent id": 9884, "Python id": 9889, "finished": true, "Ev Idx": 12789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912872.155, "dur": 6.261,
    "args": {
      "Python parent id": 9889, "Python id": 9890, "finished": true, "Ev Idx": 12790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912872.285, "dur": 4.137,
    "args": {
      "Python parent id": 9890, "Python id": 9891, "finished": true, "Ev Idx": 12791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912872.539, "dur": 3.781,
    "args": {
      "Python parent id": 9891, "Python id": 9892, "finished": true, "Ev Idx": 12792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912872.676, "dur": 0.323,
    "args": {
      "Python parent id": 9892, "Python id": 9893, "finished": true, "Ev Idx": 12793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912872.838, "dur": 0.089,
    "args": {
      "Python parent id": 9893, "Python id": 9894, "finished": true, "Ev Idx": 12794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912873.210, "dur": 2.715,
    "args": {
      "Python parent id": 9892, "Python id": 9895, "finished": true, "Ev Idx": 12795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912873.425, "dur": 2.367,
    "args": {
      "Python parent id": 9895, "Python id": 9896, "finished": true, "Ev Idx": 12796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912873.619, "dur": 2.107,
    "args": {
      "Python parent id": 9896, "Python id": 9897, "finished": true, "Ev Idx": 12797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912873.855, "dur": 1.515,
    "args": {
      "Python parent id": 9897, "Python id": 9898, "finished": true, "Ev Idx": 12798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912874.099, "dur": 0.645,
    "args": {
      "Python parent id": 9898, "Python id": 9899, "finished": true, "Ev Idx": 12799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912874.228, "dur": 0.066,
    "args": {
      "Python parent id": 9899, "Python id": 9900, "finished": true, "Ev Idx": 12800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912874.514, "dur": 0.167,
    "args": {
      "Python parent id": 9899, "Python id": 9901, "finished": true, "Ev Idx": 12801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912876.123, "dur": 0.164,
    "args": {
      "Python parent id": 9892, "Python id": 9902, "finished": true, "Ev Idx": 12802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912876.659, "dur": 0.110,
    "args": {
      "Python parent id": 9890, "Python id": 9903, "finished": true, "Ev Idx": 12803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.049, "dur": 1.321,
    "args": {
      "Python parent id": 9890, "Python id": 9904, "finished": true, "Ev Idx": 12804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.216, "dur": 1.113,
    "args": {
      "Python parent id": 9904, "Python id": 9905, "finished": true, "Ev Idx": 12805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.350, "dur": 0.481,
    "args": {
      "Python parent id": 9905, "Python id": 9906, "finished": true, "Ev Idx": 12806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.505, "dur": 0.238,
    "args": {
      "Python parent id": 9906, "Python id": 9907, "finished": true, "Ev Idx": 12807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.615, "dur": 0.079,
    "args": {
      "Python parent id": 9907, "Python id": 9908, "finished": true, "Ev Idx": 12808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912877.960, "dur": 0.302,
    "args": {
      "Python parent id": 9905, "Python id": 9909, "finished": true, "Ev Idx": 12809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912878.832, "dur": 0.288,
    "args": {
      "Python parent id": 9874, "Python id": 9910, "finished": true, "Ev Idx": 12810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912879.804, "dur": 1.125,
    "args": {
      "Python parent id": 9874, "Python id": 9911, "finished": true, "Ev Idx": 12811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912880.359, "dur": 0.518,
    "args": {
      "Python parent id": 9911, "Python id": 9912, "finished": true, "Ev Idx": 12812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912881.457, "dur": 3.417,
    "args": {
      "Python parent id": 9873, "Python id": 9913, "finished": true, "Ev Idx": 12813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912885.565, "dur": 7.798,
    "args": {
      "Python parent id": 12, "Python id": 9914, "finished": true, "Ev Idx": 12814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740912893.825, "dur": 0.084,
    "args": {
      "Python parent id": 12, "Python id": 9915, "finished": true, "Ev Idx": 12815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912894.570, "dur": 0.206,
    "args": {
      "Python parent id": 12, "Python id": 9916, "finished": true, "Ev Idx": 12816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912895.254, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 9917, "finished": true, "Ev Idx": 12817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740912895.734, "dur": 10.147,
    "args": {
      "Python parent id": 12, "Python id": 9918, "finished": true, "Ev Idx": 12818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740912907.349, "dur": 15.438,
    "args": {
      "Python parent id": 12, "Python id": 9919, "finished": true, "Ev Idx": 12819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740912907.660, "dur": 12.146,
    "args": {
      "Python parent id": 9919, "Python id": 9920, "finished": true, "Ev Idx": 12820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912907.849, "dur": 0.493,
    "args": {
      "Python parent id": 9920, "Python id": 9921, "finished": true, "Ev Idx": 12821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912908.006, "dur": 0.258,
    "args": {
      "Python parent id": 9921, "Python id": 9922, "finished": true, "Ev Idx": 12822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912908.107, "dur": 0.091,
    "args": {
      "Python parent id": 9922, "Python id": 9923, "finished": true, "Ev Idx": 12823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912908.626, "dur": 9.190,
    "args": {
      "Python parent id": 9920, "Python id": 9924, "finished": true, "Ev Idx": 12824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912908.760, "dur": 0.104,
    "args": {
      "Python parent id": 9924, "Python id": 9925, "finished": true, "Ev Idx": 12825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912908.999, "dur": 0.059,
    "args": {
      "Python parent id": 9924, "Python id": 9926, "finished": true, "Ev Idx": 12826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912909.206, "dur": 0.056,
    "args": {
      "Python parent id": 9924, "Python id": 9927, "finished": true, "Ev Idx": 12827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912909.663, "dur": 0.043,
    "args": {
      "Python parent id": 9924, "Python id": 9928, "finished": true, "Ev Idx": 12828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912910.043, "dur": 0.069,
    "args": {
      "Python parent id": 9924, "Python id": 9929, "finished": true, "Ev Idx": 12829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912910.471, "dur": 7.259,
    "args": {
      "Python parent id": 9924, "Python id": 9930, "finished": true, "Ev Idx": 12830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912910.578, "dur": 0.065,
    "args": {
      "Python parent id": 9930, "Python id": 9931, "finished": true, "Ev Idx": 12831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912910.808, "dur": 0.067,
    "args": {
      "Python parent id": 9930, "Python id": 9932, "finished": true, "Ev Idx": 12832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912911.076, "dur": 0.080,
    "args": {
      "Python parent id": 9930, "Python id": 9933, "finished": true, "Ev Idx": 12833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740912911.508, "dur": 0.042,
    "args": {
      "Python parent id": 9930, "Python id": 9934, "finished": true, "Ev Idx": 12834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740912911.920, "dur": 5.718,
    "args": {
      "Python parent id": 9930, "Python id": 9935, "finished": true, "Ev Idx": 12835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740912912.213, "dur": 5.341,
    "args": {
      "Python parent id": 9935, "Python id": 9936, "finished": true, "Ev Idx": 12836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740912912.348, "dur": 3.407,
    "args": {
      "Python parent id": 9936, "Python id": 9937, "finished": true, "Ev Idx": 12837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740912912.577, "dur": 3.102,
    "args": {
      "Python parent id": 9937, "Python id": 9938, "finished": true, "Ev Idx": 12838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740912912.703, "dur": 0.320,
    "args": {
      "Python parent id": 9938, "Python id": 9939, "finished": true, "Ev Idx": 12839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740912912.877, "dur": 0.091,
    "args": {
      "Python parent id": 9939, "Python id": 9940, "finished": true, "Ev Idx": 12840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740912913.193, "dur": 2.134,
    "args": {
      "Python parent id": 9938, "Python id": 9941, "finished": true, "Ev Idx": 12841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740912913.411, "dur": 1.829,
    "args": {
      "Python parent id": 9941, "Python id": 9942, "finished": true, "Ev Idx": 12842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740912913.598, "dur": 1.591,
    "args": {
      "Python parent id": 9942, "Python id": 9943, "finished": true, "Ev Idx": 12843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740912913.738, "dur": 1.247,
    "args": {
      "Python parent id": 9943, "Python id": 9944, "finished": true, "Ev Idx": 12844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740912913.914, "dur": 0.520,
    "args": {
      "Python parent id": 9944, "Python id": 9945, "finished": true, "Ev Idx": 12845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740912914.010, "dur": 0.077,
    "args": {
      "Python parent id": 9945, "Python id": 9946, "finished": true, "Ev Idx": 12846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740912914.242, "dur": 0.152,
    "args": {
      "Python parent id": 9945, "Python id": 9947, "finished": true, "Ev Idx": 12847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740912915.543, "dur": 0.084,
    "args": {
      "Python parent id": 9938, "Python id": 9948, "finished": true, "Ev Idx": 12848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740912915.938, "dur": 0.137,
    "args": {
      "Python parent id": 9936, "Python id": 9949, "finished": true, "Ev Idx": 12849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740912916.359, "dur": 1.126,
    "args": {
      "Python parent id": 9936, "Python id": 9950, "finished": true, "Ev Idx": 12850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740912916.501, "dur": 0.937,
    "args": {
      "Python parent id": 9950, "Python id": 9951, "finished": true, "Ev Idx": 12851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740912916.644, "dur": 0.410,
    "args": {
      "Python parent id": 9951, "Python id": 9952, "finished": true, "Ev Idx": 12852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740912916.787, "dur": 0.206,
    "args": {
      "Python parent id": 9952, "Python id": 9953, "finished": true, "Ev Idx": 12853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740912916.882, "dur": 0.055,
    "args": {
      "Python parent id": 9953, "Python id": 9954, "finished": true, "Ev Idx": 12854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740912917.189, "dur": 0.216,
    "args": {
      "Python parent id": 9951, "Python id": 9955, "finished": true, "Ev Idx": 12855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740912917.961, "dur": 0.154,
    "args": {
      "Python parent id": 9920, "Python id": 9956, "finished": true, "Ev Idx": 12856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740912918.724, "dur": 0.931,
    "args": {
      "Python parent id": 9920, "Python id": 9957, "finished": true, "Ev Idx": 12857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740912919.146, "dur": 0.457,
    "args": {
      "Python parent id": 9957, "Python id": 9958, "finished": true, "Ev Idx": 12858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740912920.097, "dur": 2.608,
    "args": {
      "Python parent id": 9919, "Python id": 9959, "finished": true, "Ev Idx": 12859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740912923.775, "dur": 101.869,
    "args": {
      "Python parent id": 12, "Python id": 9960, "finished": true, "Ev Idx": 12860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912924.119, "dur": 0.425,
    "args": {
      "Python parent id": 9960, "Python id": 9961, "finished": true, "Ev Idx": 12861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912924.379, "dur": 0.121,
    "args": {
      "Python parent id": 9961, "Python id": 9962, "finished": true, "Ev Idx": 12862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912924.830, "dur": 0.394,
    "args": {
      "Python parent id": 9960, "Python id": 9963, "finished": true, "Ev Idx": 12863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740912925.484, "dur": 0.232,
    "args": {
      "Python parent id": 9960, "Python id": 9964, "finished": true, "Ev Idx": 12864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740912925.984, "dur": 4.763,
    "args": {
      "Python parent id": 9960, "Python id": 9965, "finished": true, "Ev Idx": 12865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740912931.440, "dur": 94.132,
    "args": {
      "Python parent id": 9960, "Python id": 9966, "finished": true, "Ev Idx": 12866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740912932.266, "dur": 93.100,
    "args": {
      "Python parent id": 9966, "Python id": 9967, "finished": true, "Ev Idx": 12867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912932.596, "dur": 92.683,
    "args": {
      "Python parent id": 9967, "Python id": 9968, "finished": true, "Ev Idx": 12868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740912940.410, "dur": 83.400,
    "args": {
      "Python parent id": 9968, "Python id": 9969, "finished": true, "Ev Idx": 12869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740912940.771, "dur": 82.929,
    "args": {
      "Python parent id": 9969, "Python id": 9970, "finished": true, "Ev Idx": 12870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740912941.426, "dur": 82.178,
    "args": {
      "Python parent id": 9970, "Python id": 9971, "finished": true, "Ev Idx": 12871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912942.006, "dur": 0.455,
    "args": {
      "Python parent id": 9971, "Python id": 9972, "finished": true, "Ev Idx": 12872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912942.299, "dur": 0.064,
    "args": {
      "Python parent id": 9972, "Python id": 9973, "finished": true, "Ev Idx": 12873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740912942.815, "dur": 0.336,
    "args": {
      "Python parent id": 9971, "Python id": 9974, "finished": true, "Ev Idx": 12874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740912943.059, "dur": 0.037,
    "args": {
      "Python parent id": 9974, "Python id": 9975, "finished": true, "Ev Idx": 12875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912943.528, "dur": 79.924,
    "args": {
      "Python parent id": 9971, "Python id": 9976, "finished": true, "Ev Idx": 12876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740912944.282, "dur": 77.745,
    "args": {
      "Python parent id": 9976, "Python id": 9977, "finished": true, "Ev Idx": 12877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740912952.364, "dur": 20.574,
    "args": {
      "Python parent id": 9977, "Python id": 9978, "finished": true, "Ev Idx": 12878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740912952.512, "dur": 0.154,
    "args": {
      "Python parent id": 9978, "Python id": 9979, "finished": true, "Ev Idx": 12879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740912953.028, "dur": 19.772,
    "args": {
      "Python parent id": 9978, "Python id": 9980, "finished": true, "Ev Idx": 12880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740912953.135, "dur": 8.481,
    "args": {
      "Python parent id": 9980, "Python id": 9981, "finished": true, "Ev Idx": 12881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912981.282, "dur": 7.781,
    "args": {
      "Python parent id": 9977, "Python id": 9982, "finished": true, "Ev Idx": 12882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740912998.063, "dur": 8.582,
    "args": {
      "Python parent id": 9977, "Python id": 9983, "finished": true, "Ev Idx": 12883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913027.002, "dur": 17.745,
    "args": {
      "Python parent id": 12, "Python id": 9984, "finished": true, "Ev Idx": 12884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913027.389, "dur": 12.923,
    "args": {
      "Python parent id": 9984, "Python id": 9985, "finished": true, "Ev Idx": 12885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913027.660, "dur": 0.531,
    "args": {
      "Python parent id": 9985, "Python id": 9986, "finished": true, "Ev Idx": 12886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913027.818, "dur": 0.303,
    "args": {
      "Python parent id": 9986, "Python id": 9987, "finished": true, "Ev Idx": 12887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913027.972, "dur": 0.091,
    "args": {
      "Python parent id": 9987, "Python id": 9988, "finished": true, "Ev Idx": 12888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913028.480, "dur": 9.439,
    "args": {
      "Python parent id": 9985, "Python id": 9989, "finished": true, "Ev Idx": 12889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913028.581, "dur": 0.112,
    "args": {
      "Python parent id": 9989, "Python id": 9990, "finished": true, "Ev Idx": 12890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913028.807, "dur": 0.059,
    "args": {
      "Python parent id": 9989, "Python id": 9991, "finished": true, "Ev Idx": 12891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913029.016, "dur": 0.057,
    "args": {
      "Python parent id": 9989, "Python id": 9992, "finished": true, "Ev Idx": 12892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913029.535, "dur": 0.066,
    "args": {
      "Python parent id": 9989, "Python id": 9993, "finished": true, "Ev Idx": 12893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913029.909, "dur": 0.068,
    "args": {
      "Python parent id": 9989, "Python id": 9994, "finished": true, "Ev Idx": 12894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913030.361, "dur": 7.480,
    "args": {
      "Python parent id": 9989, "Python id": 9995, "finished": true, "Ev Idx": 12895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913030.464, "dur": 0.065,
    "args": {
      "Python parent id": 9995, "Python id": 9996, "finished": true, "Ev Idx": 12896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913030.685, "dur": 0.054,
    "args": {
      "Python parent id": 9995, "Python id": 9997, "finished": true, "Ev Idx": 12897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913030.924, "dur": 0.067,
    "args": {
      "Python parent id": 9995, "Python id": 9998, "finished": true, "Ev Idx": 12898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913031.358, "dur": 0.037,
    "args": {
      "Python parent id": 9995, "Python id": 9999, "finished": true, "Ev Idx": 12899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913031.690, "dur": 6.036,
    "args": {
      "Python parent id": 9995, "Python id": 10000, "finished": true, "Ev Idx": 12900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913031.958, "dur": 5.706,
    "args": {
      "Python parent id": 10000, "Python id": 10001, "finished": true, "Ev Idx": 12901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913032.083, "dur": 3.605,
    "args": {
      "Python parent id": 10001, "Python id": 10002, "finished": true, "Ev Idx": 12902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913032.300, "dur": 3.308,
    "args": {
      "Python parent id": 10002, "Python id": 10003, "finished": true, "Ev Idx": 12903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913032.430, "dur": 0.309,
    "args": {
      "Python parent id": 10003, "Python id": 10004, "finished": true, "Ev Idx": 12904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913032.591, "dur": 0.089,
    "args": {
      "Python parent id": 10004, "Python id": 10005, "finished": true, "Ev Idx": 12905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913032.903, "dur": 2.342,
    "args": {
      "Python parent id": 10003, "Python id": 10006, "finished": true, "Ev Idx": 12906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913033.113, "dur": 2.051,
    "args": {
      "Python parent id": 10006, "Python id": 10007, "finished": true, "Ev Idx": 12907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913033.340, "dur": 1.738,
    "args": {
      "Python parent id": 10007, "Python id": 10008, "finished": true, "Ev Idx": 12908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913033.514, "dur": 1.267,
    "args": {
      "Python parent id": 10008, "Python id": 10009, "finished": true, "Ev Idx": 12909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913033.756, "dur": 0.535,
    "args": {
      "Python parent id": 10009, "Python id": 10010, "finished": true, "Ev Idx": 12910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913033.859, "dur": 0.054,
    "args": {
      "Python parent id": 10010, "Python id": 10011, "finished": true, "Ev Idx": 12911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913034.044, "dur": 0.205,
    "args": {
      "Python parent id": 10010, "Python id": 10012, "finished": true, "Ev Idx": 12912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913035.438, "dur": 0.119,
    "args": {
      "Python parent id": 10003, "Python id": 10013, "finished": true, "Ev Idx": 12913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913035.859, "dur": 0.125,
    "args": {
      "Python parent id": 10001, "Python id": 10014, "finished": true, "Ev Idx": 12914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913036.303, "dur": 1.316,
    "args": {
      "Python parent id": 10001, "Python id": 10015, "finished": true, "Ev Idx": 12915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913036.490, "dur": 1.079,
    "args": {
      "Python parent id": 10015, "Python id": 10016, "finished": true, "Ev Idx": 12916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913036.652, "dur": 0.412,
    "args": {
      "Python parent id": 10016, "Python id": 10017, "finished": true, "Ev Idx": 12917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913036.791, "dur": 0.211,
    "args": {
      "Python parent id": 10017, "Python id": 10018, "finished": true, "Ev Idx": 12918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913036.889, "dur": 0.058,
    "args": {
      "Python parent id": 10018, "Python id": 10019, "finished": true, "Ev Idx": 12919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913037.212, "dur": 0.325,
    "args": {
      "Python parent id": 10016, "Python id": 10020, "finished": true, "Ev Idx": 12920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913038.050, "dur": 0.228,
    "args": {
      "Python parent id": 9985, "Python id": 10021, "finished": true, "Ev Idx": 12921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913039.007, "dur": 1.147,
    "args": {
      "Python parent id": 9985, "Python id": 10022, "finished": true, "Ev Idx": 12922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913039.592, "dur": 0.508,
    "args": {
      "Python parent id": 10022, "Python id": 10023, "finished": true, "Ev Idx": 12923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913040.600, "dur": 4.092,
    "args": {
      "Python parent id": 9984, "Python id": 10024, "finished": true, "Ev Idx": 12924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913045.351, "dur": 7.155,
    "args": {
      "Python parent id": 12, "Python id": 10025, "finished": true, "Ev Idx": 12925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913052.969, "dur": 0.113,
    "args": {
      "Python parent id": 12, "Python id": 10026, "finished": true, "Ev Idx": 12926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913053.741, "dur": 0.192,
    "args": {
      "Python parent id": 12, "Python id": 10027, "finished": true, "Ev Idx": 12927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913054.433, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 10028, "finished": true, "Ev Idx": 12928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913054.921, "dur": 11.236,
    "args": {
      "Python parent id": 12, "Python id": 10029, "finished": true, "Ev Idx": 12929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913067.442, "dur": 16.622,
    "args": {
      "Python parent id": 12, "Python id": 10030, "finished": true, "Ev Idx": 12930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913067.752, "dur": 11.450,
    "args": {
      "Python parent id": 10030, "Python id": 10031, "finished": true, "Ev Idx": 12931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913067.904, "dur": 0.501,
    "args": {
      "Python parent id": 10031, "Python id": 10032, "finished": true, "Ev Idx": 12932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913068.102, "dur": 0.224,
    "args": {
      "Python parent id": 10032, "Python id": 10033, "finished": true, "Ev Idx": 12933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913068.191, "dur": 0.098,
    "args": {
      "Python parent id": 10033, "Python id": 10034, "finished": true, "Ev Idx": 12934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913068.686, "dur": 8.556,
    "args": {
      "Python parent id": 10031, "Python id": 10035, "finished": true, "Ev Idx": 12935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913068.791, "dur": 0.066,
    "args": {
      "Python parent id": 10035, "Python id": 10036, "finished": true, "Ev Idx": 12936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913068.986, "dur": 0.064,
    "args": {
      "Python parent id": 10035, "Python id": 10037, "finished": true, "Ev Idx": 12937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913069.193, "dur": 0.059,
    "args": {
      "Python parent id": 10035, "Python id": 10038, "finished": true, "Ev Idx": 12938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913069.721, "dur": 0.040,
    "args": {
      "Python parent id": 10035, "Python id": 10039, "finished": true, "Ev Idx": 12939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913070.035, "dur": 0.072,
    "args": {
      "Python parent id": 10035, "Python id": 10040, "finished": true, "Ev Idx": 12940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913070.467, "dur": 6.718,
    "args": {
      "Python parent id": 10035, "Python id": 10041, "finished": true, "Ev Idx": 12941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913070.567, "dur": 0.067,
    "args": {
      "Python parent id": 10041, "Python id": 10042, "finished": true, "Ev Idx": 12942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913070.783, "dur": 0.064,
    "args": {
      "Python parent id": 10041, "Python id": 10043, "finished": true, "Ev Idx": 12943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913071.036, "dur": 0.067,
    "args": {
      "Python parent id": 10041, "Python id": 10044, "finished": true, "Ev Idx": 12944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913071.416, "dur": 0.034,
    "args": {
      "Python parent id": 10041, "Python id": 10045, "finished": true, "Ev Idx": 12945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913071.801, "dur": 5.300,
    "args": {
      "Python parent id": 10041, "Python id": 10046, "finished": true, "Ev Idx": 12946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913071.986, "dur": 5.057,
    "args": {
      "Python parent id": 10046, "Python id": 10047, "finished": true, "Ev Idx": 12947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913072.126, "dur": 3.206,
    "args": {
      "Python parent id": 10047, "Python id": 10048, "finished": true, "Ev Idx": 12948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913072.318, "dur": 2.967,
    "args": {
      "Python parent id": 10048, "Python id": 10049, "finished": true, "Ev Idx": 12949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913072.452, "dur": 0.256,
    "args": {
      "Python parent id": 10049, "Python id": 10050, "finished": true, "Ev Idx": 12950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913072.582, "dur": 0.069,
    "args": {
      "Python parent id": 10050, "Python id": 10051, "finished": true, "Ev Idx": 12951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913072.941, "dur": 2.015,
    "args": {
      "Python parent id": 10049, "Python id": 10052, "finished": true, "Ev Idx": 12952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.144, "dur": 1.750,
    "args": {
      "Python parent id": 10052, "Python id": 10053, "finished": true, "Ev Idx": 12953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.332, "dur": 1.513,
    "args": {
      "Python parent id": 10053, "Python id": 10054, "finished": true, "Ev Idx": 12954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.455, "dur": 1.219,
    "args": {
      "Python parent id": 10054, "Python id": 10055, "finished": true, "Ev Idx": 12955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.648, "dur": 0.568,
    "args": {
      "Python parent id": 10055, "Python id": 10056, "finished": true, "Ev Idx": 12956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.771, "dur": 0.048,
    "args": {
      "Python parent id": 10056, "Python id": 10057, "finished": true, "Ev Idx": 12957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913073.980, "dur": 0.197,
    "args": {
      "Python parent id": 10056, "Python id": 10058, "finished": true, "Ev Idx": 12958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913075.141, "dur": 0.107,
    "args": {
      "Python parent id": 10049, "Python id": 10059, "finished": true, "Ev Idx": 12959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913075.483, "dur": 0.079,
    "args": {
      "Python parent id": 10047, "Python id": 10060, "finished": true, "Ev Idx": 12960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913075.844, "dur": 1.152,
    "args": {
      "Python parent id": 10047, "Python id": 10061, "finished": true, "Ev Idx": 12961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913075.980, "dur": 0.972,
    "args": {
      "Python parent id": 10061, "Python id": 10062, "finished": true, "Ev Idx": 12962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913076.126, "dur": 0.402,
    "args": {
      "Python parent id": 10062, "Python id": 10063, "finished": true, "Ev Idx": 12963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913076.264, "dur": 0.202,
    "args": {
      "Python parent id": 10063, "Python id": 10064, "finished": true, "Ev Idx": 12964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913076.355, "dur": 0.082,
    "args": {
      "Python parent id": 10064, "Python id": 10065, "finished": true, "Ev Idx": 12965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913076.675, "dur": 0.248,
    "args": {
      "Python parent id": 10062, "Python id": 10066, "finished": true, "Ev Idx": 12966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913077.389, "dur": 0.154,
    "args": {
      "Python parent id": 10031, "Python id": 10067, "finished": true, "Ev Idx": 12967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913078.204, "dur": 0.845,
    "args": {
      "Python parent id": 10031, "Python id": 10068, "finished": true, "Ev Idx": 12968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913078.560, "dur": 0.439,
    "args": {
      "Python parent id": 10068, "Python id": 10069, "finished": true, "Ev Idx": 12969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913079.486, "dur": 4.523,
    "args": {
      "Python parent id": 10030, "Python id": 10070, "finished": true, "Ev Idx": 12970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913085.007, "dur": 91.015,
    "args": {
      "Python parent id": 12, "Python id": 10071, "finished": true, "Ev Idx": 12971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913085.347, "dur": 0.483,
    "args": {
      "Python parent id": 10071, "Python id": 10072, "finished": true, "Ev Idx": 12972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913085.680, "dur": 0.097,
    "args": {
      "Python parent id": 10072, "Python id": 10073, "finished": true, "Ev Idx": 12973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913086.161, "dur": 0.391,
    "args": {
      "Python parent id": 10071, "Python id": 10074, "finished": true, "Ev Idx": 12974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913086.761, "dur": 0.290,
    "args": {
      "Python parent id": 10071, "Python id": 10075, "finished": true, "Ev Idx": 12975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913087.293, "dur": 1.569,
    "args": {
      "Python parent id": 10071, "Python id": 10076, "finished": true, "Ev Idx": 12976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913089.465, "dur": 86.454,
    "args": {
      "Python parent id": 10071, "Python id": 10077, "finished": true, "Ev Idx": 12977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913090.241, "dur": 85.436,
    "args": {
      "Python parent id": 10077, "Python id": 10078, "finished": true, "Ev Idx": 12978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913090.593, "dur": 85.027,
    "args": {
      "Python parent id": 10078, "Python id": 10079, "finished": true, "Ev Idx": 12979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913095.840, "dur": 78.425,
    "args": {
      "Python parent id": 10079, "Python id": 10080, "finished": true, "Ev Idx": 12980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913096.237, "dur": 77.947,
    "args": {
      "Python parent id": 10080, "Python id": 10081, "finished": true, "Ev Idx": 12981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913096.801, "dur": 77.253,
    "args": {
      "Python parent id": 10081, "Python id": 10082, "finished": true, "Ev Idx": 12982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913097.324, "dur": 0.406,
    "args": {
      "Python parent id": 10082, "Python id": 10083, "finished": true, "Ev Idx": 12983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913097.602, "dur": 0.076,
    "args": {
      "Python parent id": 10083, "Python id": 10084, "finished": true, "Ev Idx": 12984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913098.018, "dur": 0.318,
    "args": {
      "Python parent id": 10082, "Python id": 10085, "finished": true, "Ev Idx": 12985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913098.242, "dur": 0.039,
    "args": {
      "Python parent id": 10085, "Python id": 10086, "finished": true, "Ev Idx": 12986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913098.723, "dur": 75.198,
    "args": {
      "Python parent id": 10082, "Python id": 10087, "finished": true, "Ev Idx": 12987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913099.364, "dur": 73.258,
    "args": {
      "Python parent id": 10087, "Python id": 10088, "finished": true, "Ev Idx": 12988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913106.785, "dur": 18.068,
    "args": {
      "Python parent id": 10088, "Python id": 10089, "finished": true, "Ev Idx": 12989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913106.981, "dur": 0.141,
    "args": {
      "Python parent id": 10089, "Python id": 10090, "finished": true, "Ev Idx": 12990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913107.472, "dur": 17.253,
    "args": {
      "Python parent id": 10089, "Python id": 10091, "finished": true, "Ev Idx": 12991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913107.580, "dur": 7.389,
    "args": {
      "Python parent id": 10091, "Python id": 10092, "finished": true, "Ev Idx": 12992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913132.632, "dur": 6.954,
    "args": {
      "Python parent id": 10088, "Python id": 10093, "finished": true, "Ev Idx": 12993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913148.329, "dur": 9.018,
    "args": {
      "Python parent id": 10088, "Python id": 10094, "finished": true, "Ev Idx": 12994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913177.303, "dur": 17.133,
    "args": {
      "Python parent id": 12, "Python id": 10095, "finished": true, "Ev Idx": 12995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913177.648, "dur": 12.820,
    "args": {
      "Python parent id": 10095, "Python id": 10096, "finished": true, "Ev Idx": 12996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913177.853, "dur": 0.450,
    "args": {
      "Python parent id": 10096, "Python id": 10097, "finished": true, "Ev Idx": 12997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913178.038, "dur": 0.196,
    "args": {
      "Python parent id": 10097, "Python id": 10098, "finished": true, "Ev Idx": 12998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913178.127, "dur": 0.069,
    "args": {
      "Python parent id": 10098, "Python id": 10099, "finished": true, "Ev Idx": 12999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913178.658, "dur": 9.432,
    "args": {
      "Python parent id": 10096, "Python id": 10100, "finished": true, "Ev Idx": 13000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913178.775, "dur": 0.065,
    "args": {
      "Python parent id": 10100, "Python id": 10101, "finished": true, "Ev Idx": 13001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913178.960, "dur": 0.061,
    "args": {
      "Python parent id": 10100, "Python id": 10102, "finished": true, "Ev Idx": 13002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913179.205, "dur": 0.065,
    "args": {
      "Python parent id": 10100, "Python id": 10103, "finished": true, "Ev Idx": 13003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913179.713, "dur": 0.048,
    "args": {
      "Python parent id": 10100, "Python id": 10104, "finished": true, "Ev Idx": 13004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913180.123, "dur": 0.079,
    "args": {
      "Python parent id": 10100, "Python id": 10105, "finished": true, "Ev Idx": 13005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913180.550, "dur": 7.486,
    "args": {
      "Python parent id": 10100, "Python id": 10106, "finished": true, "Ev Idx": 13006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913180.667, "dur": 0.064,
    "args": {
      "Python parent id": 10106, "Python id": 10107, "finished": true, "Ev Idx": 13007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913180.901, "dur": 0.065,
    "args": {
      "Python parent id": 10106, "Python id": 10108, "finished": true, "Ev Idx": 13008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913181.171, "dur": 0.072,
    "args": {
      "Python parent id": 10106, "Python id": 10109, "finished": true, "Ev Idx": 13009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913181.636, "dur": 0.042,
    "args": {
      "Python parent id": 10106, "Python id": 10110, "finished": true, "Ev Idx": 13010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.042, "dur": 5.896,
    "args": {
      "Python parent id": 10106, "Python id": 10111, "finished": true, "Ev Idx": 13011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.378, "dur": 5.473,
    "args": {
      "Python parent id": 10111, "Python id": 10112, "finished": true, "Ev Idx": 13012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.533, "dur": 3.527,
    "args": {
      "Python parent id": 10112, "Python id": 10113, "finished": true, "Ev Idx": 13013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.725, "dur": 3.290,
    "args": {
      "Python parent id": 10113, "Python id": 10114, "finished": true, "Ev Idx": 13014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.858, "dur": 0.272,
    "args": {
      "Python parent id": 10114, "Python id": 10115, "finished": true, "Ev Idx": 13015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913182.998, "dur": 0.069,
    "args": {
      "Python parent id": 10115, "Python id": 10116, "finished": true, "Ev Idx": 13016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913183.299, "dur": 2.302,
    "args": {
      "Python parent id": 10114, "Python id": 10117, "finished": true, "Ev Idx": 13017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913183.505, "dur": 2.014,
    "args": {
      "Python parent id": 10117, "Python id": 10118, "finished": true, "Ev Idx": 13018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913183.742, "dur": 1.731,
    "args": {
      "Python parent id": 10118, "Python id": 10119, "finished": true, "Ev Idx": 13019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913183.928, "dur": 1.308,
    "args": {
      "Python parent id": 10119, "Python id": 10120, "finished": true, "Ev Idx": 13020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913184.189, "dur": 0.554,
    "args": {
      "Python parent id": 10120, "Python id": 10121, "finished": true, "Ev Idx": 13021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913184.276, "dur": 0.070,
    "args": {
      "Python parent id": 10121, "Python id": 10122, "finished": true, "Ev Idx": 13022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913184.545, "dur": 0.160,
    "args": {
      "Python parent id": 10121, "Python id": 10123, "finished": true, "Ev Idx": 13023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913185.802, "dur": 0.179,
    "args": {
      "Python parent id": 10114, "Python id": 10124, "finished": true, "Ev Idx": 13024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913186.237, "dur": 0.104,
    "args": {
      "Python parent id": 10112, "Python id": 10125, "finished": true, "Ev Idx": 13025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913186.588, "dur": 1.218,
    "args": {
      "Python parent id": 10112, "Python id": 10126, "finished": true, "Ev Idx": 13026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913186.759, "dur": 1.001,
    "args": {
      "Python parent id": 10126, "Python id": 10127, "finished": true, "Ev Idx": 13027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913186.903, "dur": 0.430,
    "args": {
      "Python parent id": 10127, "Python id": 10128, "finished": true, "Ev Idx": 13028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913187.069, "dur": 0.206,
    "args": {
      "Python parent id": 10128, "Python id": 10129, "finished": true, "Ev Idx": 13029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913187.159, "dur": 0.063,
    "args": {
      "Python parent id": 10129, "Python id": 10130, "finished": true, "Ev Idx": 13030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913187.466, "dur": 0.259,
    "args": {
      "Python parent id": 10127, "Python id": 10131, "finished": true, "Ev Idx": 13031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913188.243, "dur": 0.300,
    "args": {
      "Python parent id": 10096, "Python id": 10132, "finished": true, "Ev Idx": 13032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913189.302, "dur": 1.033,
    "args": {
      "Python parent id": 10096, "Python id": 10133, "finished": true, "Ev Idx": 13033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913189.789, "dur": 0.496,
    "args": {
      "Python parent id": 10133, "Python id": 10134, "finished": true, "Ev Idx": 13034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913190.776, "dur": 3.565,
    "args": {
      "Python parent id": 10095, "Python id": 10135, "finished": true, "Ev Idx": 13035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913195.048, "dur": 8.315,
    "args": {
      "Python parent id": 12, "Python id": 10136, "finished": true, "Ev Idx": 13036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913203.825, "dur": 0.125,
    "args": {
      "Python parent id": 12, "Python id": 10137, "finished": true, "Ev Idx": 13037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913204.626, "dur": 0.212,
    "args": {
      "Python parent id": 12, "Python id": 10138, "finished": true, "Ev Idx": 13038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913205.341, "dur": 0.068,
    "args": {
      "Python parent id": 12, "Python id": 10139, "finished": true, "Ev Idx": 13039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913205.808, "dur": 9.482,
    "args": {
      "Python parent id": 12, "Python id": 10140, "finished": true, "Ev Idx": 13040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913216.611, "dur": 16.248,
    "args": {
      "Python parent id": 12, "Python id": 10141, "finished": true, "Ev Idx": 13041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913216.919, "dur": 11.636,
    "args": {
      "Python parent id": 10141, "Python id": 10142, "finished": true, "Ev Idx": 13042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913217.065, "dur": 0.465,
    "args": {
      "Python parent id": 10142, "Python id": 10143, "finished": true, "Ev Idx": 13043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913217.234, "dur": 0.232,
    "args": {
      "Python parent id": 10143, "Python id": 10144, "finished": true, "Ev Idx": 13044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913217.331, "dur": 0.099,
    "args": {
      "Python parent id": 10144, "Python id": 10145, "finished": true, "Ev Idx": 13045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913217.819, "dur": 8.760,
    "args": {
      "Python parent id": 10142, "Python id": 10146, "finished": true, "Ev Idx": 13046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913217.934, "dur": 0.068,
    "args": {
      "Python parent id": 10146, "Python id": 10147, "finished": true, "Ev Idx": 13047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913218.141, "dur": 0.058,
    "args": {
      "Python parent id": 10146, "Python id": 10148, "finished": true, "Ev Idx": 13048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913218.342, "dur": 0.062,
    "args": {
      "Python parent id": 10146, "Python id": 10149, "finished": true, "Ev Idx": 13049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913218.819, "dur": 0.037,
    "args": {
      "Python parent id": 10146, "Python id": 10150, "finished": true, "Ev Idx": 13050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913219.135, "dur": 0.069,
    "args": {
      "Python parent id": 10146, "Python id": 10151, "finished": true, "Ev Idx": 13051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913219.691, "dur": 6.827,
    "args": {
      "Python parent id": 10146, "Python id": 10152, "finished": true, "Ev Idx": 13052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913219.788, "dur": 0.063,
    "args": {
      "Python parent id": 10152, "Python id": 10153, "finished": true, "Ev Idx": 13053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913219.995, "dur": 0.061,
    "args": {
      "Python parent id": 10152, "Python id": 10154, "finished": true, "Ev Idx": 13054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913220.246, "dur": 0.071,
    "args": {
      "Python parent id": 10152, "Python id": 10155, "finished": true, "Ev Idx": 13055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913220.678, "dur": 0.044,
    "args": {
      "Python parent id": 10152, "Python id": 10156, "finished": true, "Ev Idx": 13056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.048, "dur": 5.375,
    "args": {
      "Python parent id": 10152, "Python id": 10157, "finished": true, "Ev Idx": 13057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.311, "dur": 5.048,
    "args": {
      "Python parent id": 10157, "Python id": 10158, "finished": true, "Ev Idx": 13058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.440, "dur": 3.122,
    "args": {
      "Python parent id": 10158, "Python id": 10159, "finished": true, "Ev Idx": 13059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.652, "dur": 2.866,
    "args": {
      "Python parent id": 10159, "Python id": 10160, "finished": true, "Ev Idx": 13060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.836, "dur": 0.298,
    "args": {
      "Python parent id": 10160, "Python id": 10161, "finished": true, "Ev Idx": 13061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913221.984, "dur": 0.066,
    "args": {
      "Python parent id": 10161, "Python id": 10162, "finished": true, "Ev Idx": 13062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913222.300, "dur": 1.927,
    "args": {
      "Python parent id": 10160, "Python id": 10163, "finished": true, "Ev Idx": 13063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913222.483, "dur": 1.666,
    "args": {
      "Python parent id": 10163, "Python id": 10164, "finished": true, "Ev Idx": 13064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913222.664, "dur": 1.436,
    "args": {
      "Python parent id": 10164, "Python id": 10165, "finished": true, "Ev Idx": 13065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913222.783, "dur": 1.108,
    "args": {
      "Python parent id": 10165, "Python id": 10166, "finished": true, "Ev Idx": 13066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913222.942, "dur": 0.496,
    "args": {
      "Python parent id": 10166, "Python id": 10167, "finished": true, "Ev Idx": 13067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913223.024, "dur": 0.070,
    "args": {
      "Python parent id": 10167, "Python id": 10168, "finished": true, "Ev Idx": 13068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913223.268, "dur": 0.122,
    "args": {
      "Python parent id": 10167, "Python id": 10169, "finished": true, "Ev Idx": 13069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913224.414, "dur": 0.066,
    "args": {
      "Python parent id": 10160, "Python id": 10170, "finished": true, "Ev Idx": 13070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913224.760, "dur": 0.076,
    "args": {
      "Python parent id": 10158, "Python id": 10171, "finished": true, "Ev Idx": 13071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.138, "dur": 1.173,
    "args": {
      "Python parent id": 10158, "Python id": 10172, "finished": true, "Ev Idx": 13072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.294, "dur": 0.970,
    "args": {
      "Python parent id": 10172, "Python id": 10173, "finished": true, "Ev Idx": 13073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.431, "dur": 0.391,
    "args": {
      "Python parent id": 10173, "Python id": 10174, "finished": true, "Ev Idx": 13074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.570, "dur": 0.184,
    "args": {
      "Python parent id": 10174, "Python id": 10175, "finished": true, "Ev Idx": 13075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.663, "dur": 0.058,
    "args": {
      "Python parent id": 10175, "Python id": 10176, "finished": true, "Ev Idx": 13076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913225.953, "dur": 0.278,
    "args": {
      "Python parent id": 10173, "Python id": 10177, "finished": true, "Ev Idx": 13077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913226.707, "dur": 0.178,
    "args": {
      "Python parent id": 10142, "Python id": 10178, "finished": true, "Ev Idx": 13078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913227.516, "dur": 0.890,
    "args": {
      "Python parent id": 10142, "Python id": 10179, "finished": true, "Ev Idx": 13079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913227.904, "dur": 0.424,
    "args": {
      "Python parent id": 10179, "Python id": 10180, "finished": true, "Ev Idx": 13080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913228.872, "dur": 3.894,
    "args": {
      "Python parent id": 10141, "Python id": 10181, "finished": true, "Ev Idx": 13081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913233.811, "dur": 92.570,
    "args": {
      "Python parent id": 12, "Python id": 10182, "finished": true, "Ev Idx": 13082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913234.193, "dur": 0.499,
    "args": {
      "Python parent id": 10182, "Python id": 10183, "finished": true, "Ev Idx": 13083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913234.502, "dur": 0.145,
    "args": {
      "Python parent id": 10183, "Python id": 10184, "finished": true, "Ev Idx": 13084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913235.057, "dur": 0.400,
    "args": {
      "Python parent id": 10182, "Python id": 10185, "finished": true, "Ev Idx": 13085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913235.662, "dur": 0.286,
    "args": {
      "Python parent id": 10182, "Python id": 10186, "finished": true, "Ev Idx": 13086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913236.160, "dur": 1.603,
    "args": {
      "Python parent id": 10182, "Python id": 10187, "finished": true, "Ev Idx": 13087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913238.340, "dur": 87.965,
    "args": {
      "Python parent id": 10182, "Python id": 10188, "finished": true, "Ev Idx": 13088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913239.252, "dur": 86.810,
    "args": {
      "Python parent id": 10188, "Python id": 10189, "finished": true, "Ev Idx": 13089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913239.628, "dur": 86.353,
    "args": {
      "Python parent id": 10189, "Python id": 10190, "finished": true, "Ev Idx": 13090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913244.870, "dur": 79.778,
    "args": {
      "Python parent id": 10190, "Python id": 10191, "finished": true, "Ev Idx": 13091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913245.191, "dur": 79.403,
    "args": {
      "Python parent id": 10191, "Python id": 10192, "finished": true, "Ev Idx": 13092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913245.823, "dur": 78.613,
    "args": {
      "Python parent id": 10192, "Python id": 10193, "finished": true, "Ev Idx": 13093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913246.343, "dur": 0.422,
    "args": {
      "Python parent id": 10193, "Python id": 10194, "finished": true, "Ev Idx": 13094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913246.633, "dur": 0.087,
    "args": {
      "Python parent id": 10194, "Python id": 10195, "finished": true, "Ev Idx": 13095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913247.047, "dur": 0.345,
    "args": {
      "Python parent id": 10193, "Python id": 10196, "finished": true, "Ev Idx": 13096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913247.298, "dur": 0.043,
    "args": {
      "Python parent id": 10196, "Python id": 10197, "finished": true, "Ev Idx": 13097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913247.719, "dur": 76.590,
    "args": {
      "Python parent id": 10193, "Python id": 10198, "finished": true, "Ev Idx": 13098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913248.444, "dur": 74.487,
    "args": {
      "Python parent id": 10198, "Python id": 10199, "finished": true, "Ev Idx": 13099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913255.628, "dur": 18.218,
    "args": {
      "Python parent id": 10199, "Python id": 10200, "finished": true, "Ev Idx": 13100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913255.761, "dur": 0.146,
    "args": {
      "Python parent id": 10200, "Python id": 10201, "finished": true, "Ev Idx": 13101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913256.223, "dur": 17.504,
    "args": {
      "Python parent id": 10200, "Python id": 10202, "finished": true, "Ev Idx": 13102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913256.382, "dur": 7.621,
    "args": {
      "Python parent id": 10202, "Python id": 10203, "finished": true, "Ev Idx": 13103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913281.000, "dur": 7.261,
    "args": {
      "Python parent id": 10199, "Python id": 10204, "finished": true, "Ev Idx": 13104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913297.943, "dur": 8.630,
    "args": {
      "Python parent id": 10199, "Python id": 10205, "finished": true, "Ev Idx": 13105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913327.616, "dur": 16.456,
    "args": {
      "Python parent id": 12, "Python id": 10206, "finished": true, "Ev Idx": 13106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913328.045, "dur": 12.326,
    "args": {
      "Python parent id": 10206, "Python id": 10207, "finished": true, "Ev Idx": 13107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913328.254, "dur": 0.455,
    "args": {
      "Python parent id": 10207, "Python id": 10208, "finished": true, "Ev Idx": 13108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913328.400, "dur": 0.240,
    "args": {
      "Python parent id": 10208, "Python id": 10209, "finished": true, "Ev Idx": 13109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913328.514, "dur": 0.091,
    "args": {
      "Python parent id": 10209, "Python id": 10210, "finished": true, "Ev Idx": 13110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913329.034, "dur": 9.167,
    "args": {
      "Python parent id": 10207, "Python id": 10211, "finished": true, "Ev Idx": 13111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913329.148, "dur": 0.103,
    "args": {
      "Python parent id": 10211, "Python id": 10212, "finished": true, "Ev Idx": 13112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913329.362, "dur": 0.067,
    "args": {
      "Python parent id": 10211, "Python id": 10213, "finished": true, "Ev Idx": 13113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913329.576, "dur": 0.060,
    "args": {
      "Python parent id": 10211, "Python id": 10214, "finished": true, "Ev Idx": 13114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913330.118, "dur": 0.045,
    "args": {
      "Python parent id": 10211, "Python id": 10215, "finished": true, "Ev Idx": 13115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913330.531, "dur": 0.075,
    "args": {
      "Python parent id": 10211, "Python id": 10216, "finished": true, "Ev Idx": 13116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913331.042, "dur": 7.102,
    "args": {
      "Python parent id": 10211, "Python id": 10217, "finished": true, "Ev Idx": 13117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913331.155, "dur": 0.063,
    "args": {
      "Python parent id": 10217, "Python id": 10218, "finished": true, "Ev Idx": 13118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913331.373, "dur": 0.056,
    "args": {
      "Python parent id": 10217, "Python id": 10219, "finished": true, "Ev Idx": 13119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913331.624, "dur": 0.069,
    "args": {
      "Python parent id": 10217, "Python id": 10220, "finished": true, "Ev Idx": 13120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913332.014, "dur": 0.037,
    "args": {
      "Python parent id": 10217, "Python id": 10221, "finished": true, "Ev Idx": 13121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913332.389, "dur": 5.664,
    "args": {
      "Python parent id": 10217, "Python id": 10222, "finished": true, "Ev Idx": 13122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913332.676, "dur": 5.294,
    "args": {
      "Python parent id": 10222, "Python id": 10223, "finished": true, "Ev Idx": 13123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913332.811, "dur": 3.350,
    "args": {
      "Python parent id": 10223, "Python id": 10224, "finished": true, "Ev Idx": 13124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913333.043, "dur": 3.077,
    "args": {
      "Python parent id": 10224, "Python id": 10225, "finished": true, "Ev Idx": 13125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913333.189, "dur": 0.254,
    "args": {
      "Python parent id": 10225, "Python id": 10226, "finished": true, "Ev Idx": 13126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913333.312, "dur": 0.070,
    "args": {
      "Python parent id": 10226, "Python id": 10227, "finished": true, "Ev Idx": 13127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913333.592, "dur": 2.120,
    "args": {
      "Python parent id": 10225, "Python id": 10228, "finished": true, "Ev Idx": 13128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913333.801, "dur": 1.817,
    "args": {
      "Python parent id": 10228, "Python id": 10229, "finished": true, "Ev Idx": 13129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913334.001, "dur": 1.571,
    "args": {
      "Python parent id": 10229, "Python id": 10230, "finished": true, "Ev Idx": 13130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913334.187, "dur": 1.183,
    "args": {
      "Python parent id": 10230, "Python id": 10231, "finished": true, "Ev Idx": 13131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913334.389, "dur": 0.497,
    "args": {
      "Python parent id": 10231, "Python id": 10232, "finished": true, "Ev Idx": 13132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913334.472, "dur": 0.076,
    "args": {
      "Python parent id": 10232, "Python id": 10233, "finished": true, "Ev Idx": 13133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913334.709, "dur": 0.123,
    "args": {
      "Python parent id": 10232, "Python id": 10234, "finished": true, "Ev Idx": 13134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913335.927, "dur": 0.150,
    "args": {
      "Python parent id": 10225, "Python id": 10235, "finished": true, "Ev Idx": 13135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913336.362, "dur": 0.073,
    "args": {
      "Python parent id": 10223, "Python id": 10236, "finished": true, "Ev Idx": 13136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913336.705, "dur": 1.220,
    "args": {
      "Python parent id": 10223, "Python id": 10237, "finished": true, "Ev Idx": 13137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913336.890, "dur": 0.990,
    "args": {
      "Python parent id": 10237, "Python id": 10238, "finished": true, "Ev Idx": 13138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913337.047, "dur": 0.422,
    "args": {
      "Python parent id": 10238, "Python id": 10239, "finished": true, "Ev Idx": 13139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913337.190, "dur": 0.219,
    "args": {
      "Python parent id": 10239, "Python id": 10240, "finished": true, "Ev Idx": 13140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913337.321, "dur": 0.056,
    "args": {
      "Python parent id": 10240, "Python id": 10241, "finished": true, "Ev Idx": 13141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913337.599, "dur": 0.248,
    "args": {
      "Python parent id": 10238, "Python id": 10242, "finished": true, "Ev Idx": 13142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913338.325, "dur": 0.251,
    "args": {
      "Python parent id": 10207, "Python id": 10243, "finished": true, "Ev Idx": 13143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913339.196, "dur": 1.011,
    "args": {
      "Python parent id": 10207, "Python id": 10244, "finished": true, "Ev Idx": 13144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913339.623, "dur": 0.534,
    "args": {
      "Python parent id": 10244, "Python id": 10245, "finished": true, "Ev Idx": 13145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913340.729, "dur": 3.280,
    "args": {
      "Python parent id": 10206, "Python id": 10246, "finished": true, "Ev Idx": 13146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913344.688, "dur": 8.154,
    "args": {
      "Python parent id": 12, "Python id": 10247, "finished": true, "Ev Idx": 13147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913353.319, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 10248, "finished": true, "Ev Idx": 13148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913354.009, "dur": 0.282,
    "args": {
      "Python parent id": 12, "Python id": 10249, "finished": true, "Ev Idx": 13149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913354.855, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 10250, "finished": true, "Ev Idx": 13150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913355.364, "dur": 10.649,
    "args": {
      "Python parent id": 12, "Python id": 10251, "finished": true, "Ev Idx": 13151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913367.280, "dur": 15.418,
    "args": {
      "Python parent id": 12, "Python id": 10252, "finished": true, "Ev Idx": 13152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913367.586, "dur": 11.504,
    "args": {
      "Python parent id": 10252, "Python id": 10253, "finished": true, "Ev Idx": 13153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913367.736, "dur": 0.460,
    "args": {
      "Python parent id": 10253, "Python id": 10254, "finished": true, "Ev Idx": 13154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913367.884, "dur": 0.234,
    "args": {
      "Python parent id": 10254, "Python id": 10255, "finished": true, "Ev Idx": 13155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913367.985, "dur": 0.095,
    "args": {
      "Python parent id": 10255, "Python id": 10256, "finished": true, "Ev Idx": 13156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913368.491, "dur": 8.769,
    "args": {
      "Python parent id": 10253, "Python id": 10257, "finished": true, "Ev Idx": 13157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913368.595, "dur": 0.404,
    "args": {
      "Python parent id": 10257, "Python id": 10258, "finished": true, "Ev Idx": 13158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913369.113, "dur": 0.065,
    "args": {
      "Python parent id": 10257, "Python id": 10259, "finished": true, "Ev Idx": 13159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913369.319, "dur": 0.059,
    "args": {
      "Python parent id": 10257, "Python id": 10260, "finished": true, "Ev Idx": 13160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913369.737, "dur": 0.036,
    "args": {
      "Python parent id": 10257, "Python id": 10261, "finished": true, "Ev Idx": 13161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913370.119, "dur": 0.076,
    "args": {
      "Python parent id": 10257, "Python id": 10262, "finished": true, "Ev Idx": 13162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913370.588, "dur": 6.604,
    "args": {
      "Python parent id": 10257, "Python id": 10263, "finished": true, "Ev Idx": 13163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913370.682, "dur": 0.067,
    "args": {
      "Python parent id": 10263, "Python id": 10264, "finished": true, "Ev Idx": 13164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913370.877, "dur": 0.090,
    "args": {
      "Python parent id": 10263, "Python id": 10265, "finished": true, "Ev Idx": 13165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913371.138, "dur": 0.066,
    "args": {
      "Python parent id": 10263, "Python id": 10266, "finished": true, "Ev Idx": 13166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913371.543, "dur": 0.029,
    "args": {
      "Python parent id": 10263, "Python id": 10267, "finished": true, "Ev Idx": 13167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913371.936, "dur": 5.194,
    "args": {
      "Python parent id": 10263, "Python id": 10268, "finished": true, "Ev Idx": 13168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913372.164, "dur": 4.902,
    "args": {
      "Python parent id": 10268, "Python id": 10269, "finished": true, "Ev Idx": 13169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913372.292, "dur": 3.016,
    "args": {
      "Python parent id": 10269, "Python id": 10270, "finished": true, "Ev Idx": 13170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913372.478, "dur": 2.782,
    "args": {
      "Python parent id": 10270, "Python id": 10271, "finished": true, "Ev Idx": 13171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913372.673, "dur": 0.245,
    "args": {
      "Python parent id": 10271, "Python id": 10272, "finished": true, "Ev Idx": 13172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913372.799, "dur": 0.066,
    "args": {
      "Python parent id": 10272, "Python id": 10273, "finished": true, "Ev Idx": 13173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.059, "dur": 1.921,
    "args": {
      "Python parent id": 10271, "Python id": 10274, "finished": true, "Ev Idx": 13174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.224, "dur": 1.692,
    "args": {
      "Python parent id": 10274, "Python id": 10275, "finished": true, "Ev Idx": 13175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.414, "dur": 1.454,
    "args": {
      "Python parent id": 10275, "Python id": 10276, "finished": true, "Ev Idx": 13176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.544, "dur": 1.149,
    "args": {
      "Python parent id": 10276, "Python id": 10277, "finished": true, "Ev Idx": 13177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.733, "dur": 0.487,
    "args": {
      "Python parent id": 10277, "Python id": 10278, "finished": true, "Ev Idx": 13178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913373.844, "dur": 0.050,
    "args": {
      "Python parent id": 10278, "Python id": 10279, "finished": true, "Ev Idx": 13179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913374.027, "dur": 0.147,
    "args": {
      "Python parent id": 10278, "Python id": 10280, "finished": true, "Ev Idx": 13180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913375.158, "dur": 0.067,
    "args": {
      "Python parent id": 10271, "Python id": 10281, "finished": true, "Ev Idx": 13181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913375.488, "dur": 0.076,
    "args": {
      "Python parent id": 10269, "Python id": 10282, "finished": true, "Ev Idx": 13182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913375.840, "dur": 1.178,
    "args": {
      "Python parent id": 10269, "Python id": 10283, "finished": true, "Ev Idx": 13183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913376.001, "dur": 0.970,
    "args": {
      "Python parent id": 10283, "Python id": 10284, "finished": true, "Ev Idx": 13184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913376.189, "dur": 0.365,
    "args": {
      "Python parent id": 10284, "Python id": 10285, "finished": true, "Ev Idx": 13185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913376.318, "dur": 0.175,
    "args": {
      "Python parent id": 10285, "Python id": 10286, "finished": true, "Ev Idx": 13186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913376.407, "dur": 0.055,
    "args": {
      "Python parent id": 10286, "Python id": 10287, "finished": true, "Ev Idx": 13187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913376.684, "dur": 0.254,
    "args": {
      "Python parent id": 10284, "Python id": 10288, "finished": true, "Ev Idx": 13188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913377.389, "dur": 0.143,
    "args": {
      "Python parent id": 10253, "Python id": 10289, "finished": true, "Ev Idx": 13189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913378.080, "dur": 0.881,
    "args": {
      "Python parent id": 10253, "Python id": 10290, "finished": true, "Ev Idx": 13190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913378.510, "dur": 0.398,
    "args": {
      "Python parent id": 10290, "Python id": 10291, "finished": true, "Ev Idx": 13191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913379.378, "dur": 3.262,
    "args": {
      "Python parent id": 10252, "Python id": 10292, "finished": true, "Ev Idx": 13192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913383.658, "dur": 92.081,
    "args": {
      "Python parent id": 12, "Python id": 10293, "finished": true, "Ev Idx": 13193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913384.000, "dur": 0.450,
    "args": {
      "Python parent id": 10293, "Python id": 10294, "finished": true, "Ev Idx": 13194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913384.281, "dur": 0.089,
    "args": {
      "Python parent id": 10294, "Python id": 10295, "finished": true, "Ev Idx": 13195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913384.802, "dur": 0.442,
    "args": {
      "Python parent id": 10293, "Python id": 10296, "finished": true, "Ev Idx": 13196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913385.496, "dur": 0.282,
    "args": {
      "Python parent id": 10293, "Python id": 10297, "finished": true, "Ev Idx": 13197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913386.007, "dur": 1.543,
    "args": {
      "Python parent id": 10293, "Python id": 10298, "finished": true, "Ev Idx": 13198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913388.128, "dur": 87.538,
    "args": {
      "Python parent id": 10293, "Python id": 10299, "finished": true, "Ev Idx": 13199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913388.951, "dur": 86.532,
    "args": {
      "Python parent id": 10299, "Python id": 10300, "finished": true, "Ev Idx": 13200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913389.312, "dur": 86.087,
    "args": {
      "Python parent id": 10300, "Python id": 10301, "finished": true, "Ev Idx": 13201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913394.559, "dur": 79.540,
    "args": {
      "Python parent id": 10301, "Python id": 10302, "finished": true, "Ev Idx": 13202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913394.929, "dur": 79.078,
    "args": {
      "Python parent id": 10302, "Python id": 10303, "finished": true, "Ev Idx": 13203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913395.508, "dur": 78.334,
    "args": {
      "Python parent id": 10303, "Python id": 10304, "finished": true, "Ev Idx": 13204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913395.996, "dur": 0.423,
    "args": {
      "Python parent id": 10304, "Python id": 10305, "finished": true, "Ev Idx": 13205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913396.289, "dur": 0.045,
    "args": {
      "Python parent id": 10305, "Python id": 10306, "finished": true, "Ev Idx": 13206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913396.729, "dur": 0.316,
    "args": {
      "Python parent id": 10304, "Python id": 10307, "finished": true, "Ev Idx": 13207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913396.953, "dur": 0.039,
    "args": {
      "Python parent id": 10307, "Python id": 10308, "finished": true, "Ev Idx": 13208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913397.393, "dur": 76.316,
    "args": {
      "Python parent id": 10304, "Python id": 10309, "finished": true, "Ev Idx": 13209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913398.072, "dur": 74.310,
    "args": {
      "Python parent id": 10309, "Python id": 10310, "finished": true, "Ev Idx": 13210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913405.188, "dur": 18.358,
    "args": {
      "Python parent id": 10310, "Python id": 10311, "finished": true, "Ev Idx": 13211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913405.340, "dur": 0.143,
    "args": {
      "Python parent id": 10311, "Python id": 10312, "finished": true, "Ev Idx": 13212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913405.794, "dur": 17.636,
    "args": {
      "Python parent id": 10311, "Python id": 10313, "finished": true, "Ev Idx": 13213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913405.912, "dur": 8.139,
    "args": {
      "Python parent id": 10313, "Python id": 10314, "finished": true, "Ev Idx": 13214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913431.773, "dur": 7.303,
    "args": {
      "Python parent id": 10310, "Python id": 10315, "finished": true, "Ev Idx": 13215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913447.773, "dur": 8.315,
    "args": {
      "Python parent id": 10310, "Python id": 10316, "finished": true, "Ev Idx": 13216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913477.014, "dur": 17.505,
    "args": {
      "Python parent id": 12, "Python id": 10317, "finished": true, "Ev Idx": 13217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913477.451, "dur": 12.985,
    "args": {
      "Python parent id": 10317, "Python id": 10318, "finished": true, "Ev Idx": 13218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913477.692, "dur": 0.494,
    "args": {
      "Python parent id": 10318, "Python id": 10319, "finished": true, "Ev Idx": 13219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913477.859, "dur": 0.226,
    "args": {
      "Python parent id": 10319, "Python id": 10320, "finished": true, "Ev Idx": 13220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913477.973, "dur": 0.076,
    "args": {
      "Python parent id": 10320, "Python id": 10321, "finished": true, "Ev Idx": 13221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913478.543, "dur": 9.554,
    "args": {
      "Python parent id": 10318, "Python id": 10322, "finished": true, "Ev Idx": 13222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913478.669, "dur": 0.070,
    "args": {
      "Python parent id": 10322, "Python id": 10323, "finished": true, "Ev Idx": 13223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913478.849, "dur": 0.064,
    "args": {
      "Python parent id": 10322, "Python id": 10324, "finished": true, "Ev Idx": 13224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913479.062, "dur": 0.063,
    "args": {
      "Python parent id": 10322, "Python id": 10325, "finished": true, "Ev Idx": 13225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913479.602, "dur": 0.057,
    "args": {
      "Python parent id": 10322, "Python id": 10326, "finished": true, "Ev Idx": 13226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913479.977, "dur": 0.075,
    "args": {
      "Python parent id": 10322, "Python id": 10327, "finished": true, "Ev Idx": 13227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913480.377, "dur": 7.655,
    "args": {
      "Python parent id": 10322, "Python id": 10328, "finished": true, "Ev Idx": 13228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913480.505, "dur": 0.065,
    "args": {
      "Python parent id": 10328, "Python id": 10329, "finished": true, "Ev Idx": 13229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913480.758, "dur": 0.058,
    "args": {
      "Python parent id": 10328, "Python id": 10330, "finished": true, "Ev Idx": 13230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913480.999, "dur": 0.065,
    "args": {
      "Python parent id": 10328, "Python id": 10331, "finished": true, "Ev Idx": 13231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913481.384, "dur": 0.068,
    "args": {
      "Python parent id": 10328, "Python id": 10332, "finished": true, "Ev Idx": 13232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913481.818, "dur": 6.157,
    "args": {
      "Python parent id": 10328, "Python id": 10333, "finished": true, "Ev Idx": 13233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913482.164, "dur": 5.748,
    "args": {
      "Python parent id": 10333, "Python id": 10334, "finished": true, "Ev Idx": 13234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913482.290, "dur": 3.736,
    "args": {
      "Python parent id": 10334, "Python id": 10335, "finished": true, "Ev Idx": 13235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913482.485, "dur": 3.450,
    "args": {
      "Python parent id": 10335, "Python id": 10336, "finished": true, "Ev Idx": 13236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913482.661, "dur": 0.332,
    "args": {
      "Python parent id": 10336, "Python id": 10337, "finished": true, "Ev Idx": 13237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913482.830, "dur": 0.070,
    "args": {
      "Python parent id": 10337, "Python id": 10338, "finished": true, "Ev Idx": 13238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913483.139, "dur": 2.388,
    "args": {
      "Python parent id": 10336, "Python id": 10339, "finished": true, "Ev Idx": 13239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913483.395, "dur": 2.015,
    "args": {
      "Python parent id": 10339, "Python id": 10340, "finished": true, "Ev Idx": 13240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913483.583, "dur": 1.746,
    "args": {
      "Python parent id": 10340, "Python id": 10341, "finished": true, "Ev Idx": 13241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913483.795, "dur": 1.322,
    "args": {
      "Python parent id": 10341, "Python id": 10342, "finished": true, "Ev Idx": 13242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913483.985, "dur": 0.565,
    "args": {
      "Python parent id": 10342, "Python id": 10343, "finished": true, "Ev Idx": 13243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913484.094, "dur": 0.076,
    "args": {
      "Python parent id": 10343, "Python id": 10344, "finished": true, "Ev Idx": 13244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913484.321, "dur": 0.188,
    "args": {
      "Python parent id": 10343, "Python id": 10345, "finished": true, "Ev Idx": 13245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913485.701, "dur": 0.171,
    "args": {
      "Python parent id": 10336, "Python id": 10346, "finished": true, "Ev Idx": 13246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913486.190, "dur": 0.108,
    "args": {
      "Python parent id": 10334, "Python id": 10347, "finished": true, "Ev Idx": 13247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913486.590, "dur": 1.233,
    "args": {
      "Python parent id": 10334, "Python id": 10348, "finished": true, "Ev Idx": 13248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913486.789, "dur": 0.993,
    "args": {
      "Python parent id": 10348, "Python id": 10349, "finished": true, "Ev Idx": 13249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913486.935, "dur": 0.410,
    "args": {
      "Python parent id": 10349, "Python id": 10350, "finished": true, "Ev Idx": 13250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913487.109, "dur": 0.175,
    "args": {
      "Python parent id": 10350, "Python id": 10351, "finished": true, "Ev Idx": 13251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913487.192, "dur": 0.057,
    "args": {
      "Python parent id": 10351, "Python id": 10352, "finished": true, "Ev Idx": 13252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913487.471, "dur": 0.267,
    "args": {
      "Python parent id": 10349, "Python id": 10353, "finished": true, "Ev Idx": 13253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913488.246, "dur": 0.227,
    "args": {
      "Python parent id": 10318, "Python id": 10354, "finished": true, "Ev Idx": 13254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913489.178, "dur": 1.094,
    "args": {
      "Python parent id": 10318, "Python id": 10355, "finished": true, "Ev Idx": 13255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913489.678, "dur": 0.525,
    "args": {
      "Python parent id": 10355, "Python id": 10356, "finished": true, "Ev Idx": 13256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913490.803, "dur": 3.615,
    "args": {
      "Python parent id": 10317, "Python id": 10357, "finished": true, "Ev Idx": 13257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913495.127, "dur": 7.494,
    "args": {
      "Python parent id": 12, "Python id": 10358, "finished": true, "Ev Idx": 13258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913503.071, "dur": 0.112,
    "args": {
      "Python parent id": 12, "Python id": 10359, "finished": true, "Ev Idx": 13259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913503.794, "dur": 0.166,
    "args": {
      "Python parent id": 12, "Python id": 10360, "finished": true, "Ev Idx": 13260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913504.446, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 10361, "finished": true, "Ev Idx": 13261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913504.909, "dur": 22.107,
    "args": {
      "Python parent id": 12, "Python id": 10362, "finished": true, "Ev Idx": 13262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913529.466, "dur": 20.900,
    "args": {
      "Python parent id": 12, "Python id": 10363, "finished": true, "Ev Idx": 13263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913529.920, "dur": 14.416,
    "args": {
      "Python parent id": 10363, "Python id": 10364, "finished": true, "Ev Idx": 13264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913530.121, "dur": 0.691,
    "args": {
      "Python parent id": 10364, "Python id": 10365, "finished": true, "Ev Idx": 13265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913530.280, "dur": 0.415,
    "args": {
      "Python parent id": 10365, "Python id": 10366, "finished": true, "Ev Idx": 13266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913530.449, "dur": 0.188,
    "args": {
      "Python parent id": 10366, "Python id": 10367, "finished": true, "Ev Idx": 13267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913531.193, "dur": 10.443,
    "args": {
      "Python parent id": 10364, "Python id": 10368, "finished": true, "Ev Idx": 13268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913531.323, "dur": 0.068,
    "args": {
      "Python parent id": 10368, "Python id": 10369, "finished": true, "Ev Idx": 13269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913531.538, "dur": 0.068,
    "args": {
      "Python parent id": 10368, "Python id": 10370, "finished": true, "Ev Idx": 13270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913531.779, "dur": 0.057,
    "args": {
      "Python parent id": 10368, "Python id": 10371, "finished": true, "Ev Idx": 13271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913532.272, "dur": 0.061,
    "args": {
      "Python parent id": 10368, "Python id": 10372, "finished": true, "Ev Idx": 13272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913532.658, "dur": 0.079,
    "args": {
      "Python parent id": 10368, "Python id": 10373, "finished": true, "Ev Idx": 13273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913533.224, "dur": 8.318,
    "args": {
      "Python parent id": 10368, "Python id": 10374, "finished": true, "Ev Idx": 13274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913533.308, "dur": 0.061,
    "args": {
      "Python parent id": 10374, "Python id": 10375, "finished": true, "Ev Idx": 13275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913533.540, "dur": 0.066,
    "args": {
      "Python parent id": 10374, "Python id": 10376, "finished": true, "Ev Idx": 13276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913533.810, "dur": 0.064,
    "args": {
      "Python parent id": 10374, "Python id": 10377, "finished": true, "Ev Idx": 13277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913534.254, "dur": 0.068,
    "args": {
      "Python parent id": 10374, "Python id": 10378, "finished": true, "Ev Idx": 13278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913534.711, "dur": 6.690,
    "args": {
      "Python parent id": 10374, "Python id": 10379, "finished": true, "Ev Idx": 13279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913534.964, "dur": 6.332,
    "args": {
      "Python parent id": 10379, "Python id": 10380, "finished": true, "Ev Idx": 13280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913535.108, "dur": 4.060,
    "args": {
      "Python parent id": 10380, "Python id": 10381, "finished": true, "Ev Idx": 13281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913535.295, "dur": 3.827,
    "args": {
      "Python parent id": 10381, "Python id": 10382, "finished": true, "Ev Idx": 13282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913535.454, "dur": 0.301,
    "args": {
      "Python parent id": 10382, "Python id": 10383, "finished": true, "Ev Idx": 13283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913535.614, "dur": 0.073,
    "args": {
      "Python parent id": 10383, "Python id": 10384, "finished": true, "Ev Idx": 13284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913535.958, "dur": 2.799,
    "args": {
      "Python parent id": 10382, "Python id": 10385, "finished": true, "Ev Idx": 13285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913536.237, "dur": 2.389,
    "args": {
      "Python parent id": 10385, "Python id": 10386, "finished": true, "Ev Idx": 13286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913536.423, "dur": 2.135,
    "args": {
      "Python parent id": 10386, "Python id": 10387, "finished": true, "Ev Idx": 13287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913536.595, "dur": 1.601,
    "args": {
      "Python parent id": 10387, "Python id": 10388, "finished": true, "Ev Idx": 13288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913536.808, "dur": 0.588,
    "args": {
      "Python parent id": 10388, "Python id": 10389, "finished": true, "Ev Idx": 13289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913536.908, "dur": 0.050,
    "args": {
      "Python parent id": 10389, "Python id": 10390, "finished": true, "Ev Idx": 13290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913537.151, "dur": 0.201,
    "args": {
      "Python parent id": 10389, "Python id": 10391, "finished": true, "Ev Idx": 13291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913538.940, "dur": 0.136,
    "args": {
      "Python parent id": 10382, "Python id": 10392, "finished": true, "Ev Idx": 13292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913539.380, "dur": 0.118,
    "args": {
      "Python parent id": 10380, "Python id": 10393, "finished": true, "Ev Idx": 13293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913539.856, "dur": 1.373,
    "args": {
      "Python parent id": 10380, "Python id": 10394, "finished": true, "Ev Idx": 13294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913539.999, "dur": 1.187,
    "args": {
      "Python parent id": 10394, "Python id": 10395, "finished": true, "Ev Idx": 13295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913540.139, "dur": 0.415,
    "args": {
      "Python parent id": 10395, "Python id": 10396, "finished": true, "Ev Idx": 13296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913540.304, "dur": 0.192,
    "args": {
      "Python parent id": 10396, "Python id": 10397, "finished": true, "Ev Idx": 13297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913540.400, "dur": 0.059,
    "args": {
      "Python parent id": 10397, "Python id": 10398, "finished": true, "Ev Idx": 13298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913540.696, "dur": 0.445,
    "args": {
      "Python parent id": 10395, "Python id": 10399, "finished": true, "Ev Idx": 13299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913541.766, "dur": 0.313,
    "args": {
      "Python parent id": 10364, "Python id": 10400, "finished": true, "Ev Idx": 13300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913542.808, "dur": 1.358,
    "args": {
      "Python parent id": 10364, "Python id": 10401, "finished": true, "Ev Idx": 13301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913543.374, "dur": 0.684,
    "args": {
      "Python parent id": 10401, "Python id": 10402, "finished": true, "Ev Idx": 13302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913544.694, "dur": 5.576,
    "args": {
      "Python parent id": 10363, "Python id": 10403, "finished": true, "Ev Idx": 13303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913551.397, "dur": 121.478,
    "args": {
      "Python parent id": 12, "Python id": 10404, "finished": true, "Ev Idx": 13304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913551.812, "dur": 0.420,
    "args": {
      "Python parent id": 10404, "Python id": 10405, "finished": true, "Ev Idx": 13305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913552.070, "dur": 0.120,
    "args": {
      "Python parent id": 10405, "Python id": 10406, "finished": true, "Ev Idx": 13306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913552.558, "dur": 0.470,
    "args": {
      "Python parent id": 10404, "Python id": 10407, "finished": true, "Ev Idx": 13307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913553.280, "dur": 0.273,
    "args": {
      "Python parent id": 10404, "Python id": 10408, "finished": true, "Ev Idx": 13308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913553.794, "dur": 2.521,
    "args": {
      "Python parent id": 10404, "Python id": 10409, "finished": true, "Ev Idx": 13309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913556.926, "dur": 115.843,
    "args": {
      "Python parent id": 10404, "Python id": 10410, "finished": true, "Ev Idx": 13310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913557.940, "dur": 114.637,
    "args": {
      "Python parent id": 10410, "Python id": 10411, "finished": true, "Ev Idx": 13311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913558.296, "dur": 114.225,
    "args": {
      "Python parent id": 10411, "Python id": 10412, "finished": true, "Ev Idx": 13312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913565.509, "dur": 105.273,
    "args": {
      "Python parent id": 10412, "Python id": 10413, "finished": true, "Ev Idx": 13313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913565.898, "dur": 104.798,
    "args": {
      "Python parent id": 10413, "Python id": 10414, "finished": true, "Ev Idx": 13314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913566.631, "dur": 103.945,
    "args": {
      "Python parent id": 10414, "Python id": 10415, "finished": true, "Ev Idx": 13315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913567.127, "dur": 0.456,
    "args": {
      "Python parent id": 10415, "Python id": 10416, "finished": true, "Ev Idx": 13316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913567.461, "dur": 0.038,
    "args": {
      "Python parent id": 10416, "Python id": 10417, "finished": true, "Ev Idx": 13317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913567.903, "dur": 0.324,
    "args": {
      "Python parent id": 10415, "Python id": 10418, "finished": true, "Ev Idx": 13318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913568.136, "dur": 0.036,
    "args": {
      "Python parent id": 10418, "Python id": 10419, "finished": true, "Ev Idx": 13319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913568.626, "dur": 101.832,
    "args": {
      "Python parent id": 10415, "Python id": 10420, "finished": true, "Ev Idx": 13320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913569.430, "dur": 99.438,
    "args": {
      "Python parent id": 10420, "Python id": 10421, "finished": true, "Ev Idx": 13321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913577.979, "dur": 38.184,
    "args": {
      "Python parent id": 10421, "Python id": 10422, "finished": true, "Ev Idx": 13322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913578.179, "dur": 0.153,
    "args": {
      "Python parent id": 10422, "Python id": 10423, "finished": true, "Ev Idx": 13323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913578.723, "dur": 37.314,
    "args": {
      "Python parent id": 10422, "Python id": 10424, "finished": true, "Ev Idx": 13324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913578.856, "dur": 24.449,
    "args": {
      "Python parent id": 10424, "Python id": 10425, "finished": true, "Ev Idx": 13325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913625.179, "dur": 7.673,
    "args": {
      "Python parent id": 10421, "Python id": 10426, "finished": true, "Ev Idx": 13326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913642.928, "dur": 10.405,
    "args": {
      "Python parent id": 10421, "Python id": 10427, "finished": true, "Ev Idx": 13327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913674.407, "dur": 19.689,
    "args": {
      "Python parent id": 12, "Python id": 10428, "finished": true, "Ev Idx": 13328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913674.815, "dur": 14.387,
    "args": {
      "Python parent id": 10428, "Python id": 10429, "finished": true, "Ev Idx": 13329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913675.038, "dur": 0.616,
    "args": {
      "Python parent id": 10429, "Python id": 10430, "finished": true, "Ev Idx": 13330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913675.255, "dur": 0.296,
    "args": {
      "Python parent id": 10430, "Python id": 10431, "finished": true, "Ev Idx": 13331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913675.400, "dur": 0.120,
    "args": {
      "Python parent id": 10431, "Python id": 10432, "finished": true, "Ev Idx": 13332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913676.000, "dur": 10.434,
    "args": {
      "Python parent id": 10429, "Python id": 10433, "finished": true, "Ev Idx": 13333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913676.092, "dur": 0.119,
    "args": {
      "Python parent id": 10433, "Python id": 10434, "finished": true, "Ev Idx": 13334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913676.352, "dur": 0.063,
    "args": {
      "Python parent id": 10433, "Python id": 10435, "finished": true, "Ev Idx": 13335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913676.601, "dur": 0.060,
    "args": {
      "Python parent id": 10433, "Python id": 10436, "finished": true, "Ev Idx": 13336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913677.159, "dur": 0.061,
    "args": {
      "Python parent id": 10433, "Python id": 10437, "finished": true, "Ev Idx": 13337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913677.492, "dur": 0.066,
    "args": {
      "Python parent id": 10433, "Python id": 10438, "finished": true, "Ev Idx": 13338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913677.952, "dur": 8.399,
    "args": {
      "Python parent id": 10433, "Python id": 10439, "finished": true, "Ev Idx": 13339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913678.069, "dur": 0.062,
    "args": {
      "Python parent id": 10439, "Python id": 10440, "finished": true, "Ev Idx": 13340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913678.302, "dur": 0.062,
    "args": {
      "Python parent id": 10439, "Python id": 10441, "finished": true, "Ev Idx": 13341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913678.564, "dur": 0.069,
    "args": {
      "Python parent id": 10439, "Python id": 10442, "finished": true, "Ev Idx": 13342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913679.033, "dur": 0.029,
    "args": {
      "Python parent id": 10439, "Python id": 10443, "finished": true, "Ev Idx": 13343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913679.474, "dur": 6.759,
    "args": {
      "Python parent id": 10439, "Python id": 10444, "finished": true, "Ev Idx": 13344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913679.736, "dur": 6.412,
    "args": {
      "Python parent id": 10444, "Python id": 10445, "finished": true, "Ev Idx": 13345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913679.893, "dur": 4.166,
    "args": {
      "Python parent id": 10445, "Python id": 10446, "finished": true, "Ev Idx": 13346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913680.080, "dur": 3.889,
    "args": {
      "Python parent id": 10446, "Python id": 10447, "finished": true, "Ev Idx": 13347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913680.259, "dur": 0.293,
    "args": {
      "Python parent id": 10447, "Python id": 10448, "finished": true, "Ev Idx": 13348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913680.423, "dur": 0.067,
    "args": {
      "Python parent id": 10448, "Python id": 10449, "finished": true, "Ev Idx": 13349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913680.753, "dur": 2.783,
    "args": {
      "Python parent id": 10447, "Python id": 10450, "finished": true, "Ev Idx": 13350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913681.030, "dur": 2.357,
    "args": {
      "Python parent id": 10450, "Python id": 10451, "finished": true, "Ev Idx": 13351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913681.244, "dur": 2.056,
    "args": {
      "Python parent id": 10451, "Python id": 10452, "finished": true, "Ev Idx": 13352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913681.452, "dur": 1.520,
    "args": {
      "Python parent id": 10452, "Python id": 10453, "finished": true, "Ev Idx": 13353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913681.700, "dur": 0.634,
    "args": {
      "Python parent id": 10453, "Python id": 10454, "finished": true, "Ev Idx": 13354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913681.823, "dur": 0.066,
    "args": {
      "Python parent id": 10454, "Python id": 10455, "finished": true, "Ev Idx": 13355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913682.105, "dur": 0.174,
    "args": {
      "Python parent id": 10454, "Python id": 10456, "finished": true, "Ev Idx": 13356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913683.745, "dur": 0.171,
    "args": {
      "Python parent id": 10447, "Python id": 10457, "finished": true, "Ev Idx": 13357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913684.258, "dur": 0.181,
    "args": {
      "Python parent id": 10445, "Python id": 10458, "finished": true, "Ev Idx": 13358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913684.735, "dur": 1.356,
    "args": {
      "Python parent id": 10445, "Python id": 10459, "finished": true, "Ev Idx": 13359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913684.900, "dur": 1.128,
    "args": {
      "Python parent id": 10459, "Python id": 10460, "finished": true, "Ev Idx": 13360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913685.037, "dur": 0.434,
    "args": {
      "Python parent id": 10460, "Python id": 10461, "finished": true, "Ev Idx": 13361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913685.181, "dur": 0.206,
    "args": {
      "Python parent id": 10461, "Python id": 10462, "finished": true, "Ev Idx": 13362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913685.296, "dur": 0.056,
    "args": {
      "Python parent id": 10462, "Python id": 10463, "finished": true, "Ev Idx": 13363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913685.639, "dur": 0.322,
    "args": {
      "Python parent id": 10460, "Python id": 10464, "finished": true, "Ev Idx": 13364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913686.575, "dur": 0.332,
    "args": {
      "Python parent id": 10429, "Python id": 10465, "finished": true, "Ev Idx": 13365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913687.617, "dur": 1.394,
    "args": {
      "Python parent id": 10429, "Python id": 10466, "finished": true, "Ev Idx": 13366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913688.196, "dur": 0.724,
    "args": {
      "Python parent id": 10466, "Python id": 10467, "finished": true, "Ev Idx": 13367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913689.643, "dur": 4.339,
    "args": {
      "Python parent id": 10428, "Python id": 10468, "finished": true, "Ev Idx": 13368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913694.713, "dur": 8.723,
    "args": {
      "Python parent id": 12, "Python id": 10469, "finished": true, "Ev Idx": 13369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913703.941, "dur": 0.106,
    "args": {
      "Python parent id": 12, "Python id": 10470, "finished": true, "Ev Idx": 13370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913704.920, "dur": 0.229,
    "args": {
      "Python parent id": 12, "Python id": 10471, "finished": true, "Ev Idx": 13371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913705.744, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 10472, "finished": true, "Ev Idx": 13372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913706.213, "dur": 11.721,
    "args": {
      "Python parent id": 12, "Python id": 10473, "finished": true, "Ev Idx": 13373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913719.328, "dur": 15.926,
    "args": {
      "Python parent id": 12, "Python id": 10474, "finished": true, "Ev Idx": 13374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913719.658, "dur": 12.244,
    "args": {
      "Python parent id": 10474, "Python id": 10475, "finished": true, "Ev Idx": 13375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913719.836, "dur": 0.555,
    "args": {
      "Python parent id": 10475, "Python id": 10476, "finished": true, "Ev Idx": 13376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913720.014, "dur": 0.286,
    "args": {
      "Python parent id": 10476, "Python id": 10477, "finished": true, "Ev Idx": 13377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913720.134, "dur": 0.125,
    "args": {
      "Python parent id": 10477, "Python id": 10478, "finished": true, "Ev Idx": 13378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913720.773, "dur": 8.997,
    "args": {
      "Python parent id": 10475, "Python id": 10479, "finished": true, "Ev Idx": 13379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913720.865, "dur": 0.064,
    "args": {
      "Python parent id": 10479, "Python id": 10480, "finished": true, "Ev Idx": 13380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913721.071, "dur": 0.062,
    "args": {
      "Python parent id": 10479, "Python id": 10481, "finished": true, "Ev Idx": 13381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913721.346, "dur": 0.058,
    "args": {
      "Python parent id": 10479, "Python id": 10482, "finished": true, "Ev Idx": 13382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913721.733, "dur": 0.039,
    "args": {
      "Python parent id": 10479, "Python id": 10483, "finished": true, "Ev Idx": 13383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913722.054, "dur": 0.073,
    "args": {
      "Python parent id": 10479, "Python id": 10484, "finished": true, "Ev Idx": 13384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913722.494, "dur": 7.183,
    "args": {
      "Python parent id": 10479, "Python id": 10485, "finished": true, "Ev Idx": 13385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913722.610, "dur": 0.070,
    "args": {
      "Python parent id": 10485, "Python id": 10486, "finished": true, "Ev Idx": 13386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913722.836, "dur": 0.061,
    "args": {
      "Python parent id": 10485, "Python id": 10487, "finished": true, "Ev Idx": 13387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913723.109, "dur": 0.068,
    "args": {
      "Python parent id": 10485, "Python id": 10488, "finished": true, "Ev Idx": 13388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913723.554, "dur": 0.042,
    "args": {
      "Python parent id": 10485, "Python id": 10489, "finished": true, "Ev Idx": 13389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913723.989, "dur": 5.579,
    "args": {
      "Python parent id": 10485, "Python id": 10490, "finished": true, "Ev Idx": 13390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913724.264, "dur": 5.246,
    "args": {
      "Python parent id": 10490, "Python id": 10491, "finished": true, "Ev Idx": 13391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913724.419, "dur": 3.246,
    "args": {
      "Python parent id": 10491, "Python id": 10492, "finished": true, "Ev Idx": 13392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913724.602, "dur": 2.992,
    "args": {
      "Python parent id": 10492, "Python id": 10493, "finished": true, "Ev Idx": 13393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913724.727, "dur": 0.313,
    "args": {
      "Python parent id": 10493, "Python id": 10494, "finished": true, "Ev Idx": 13394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913724.898, "dur": 0.078,
    "args": {
      "Python parent id": 10494, "Python id": 10495, "finished": true, "Ev Idx": 13395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913725.181, "dur": 2.123,
    "args": {
      "Python parent id": 10493, "Python id": 10496, "finished": true, "Ev Idx": 13396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913725.368, "dur": 1.846,
    "args": {
      "Python parent id": 10496, "Python id": 10497, "finished": true, "Ev Idx": 13397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913725.540, "dur": 1.631,
    "args": {
      "Python parent id": 10497, "Python id": 10498, "finished": true, "Ev Idx": 13398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913725.705, "dur": 1.243,
    "args": {
      "Python parent id": 10498, "Python id": 10499, "finished": true, "Ev Idx": 13399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913725.912, "dur": 0.510,
    "args": {
      "Python parent id": 10499, "Python id": 10500, "finished": true, "Ev Idx": 13400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913726.009, "dur": 0.051,
    "args": {
      "Python parent id": 10500, "Python id": 10501, "finished": true, "Ev Idx": 13401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913726.250, "dur": 0.128,
    "args": {
      "Python parent id": 10500, "Python id": 10502, "finished": true, "Ev Idx": 13402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913727.486, "dur": 0.076,
    "args": {
      "Python parent id": 10493, "Python id": 10503, "finished": true, "Ev Idx": 13403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913727.836, "dur": 0.140,
    "args": {
      "Python parent id": 10491, "Python id": 10504, "finished": true, "Ev Idx": 13404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913728.270, "dur": 1.173,
    "args": {
      "Python parent id": 10491, "Python id": 10505, "finished": true, "Ev Idx": 13405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913728.412, "dur": 0.989,
    "args": {
      "Python parent id": 10505, "Python id": 10506, "finished": true, "Ev Idx": 13406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913728.581, "dur": 0.409,
    "args": {
      "Python parent id": 10506, "Python id": 10507, "finished": true, "Ev Idx": 13407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913728.750, "dur": 0.182,
    "args": {
      "Python parent id": 10507, "Python id": 10508, "finished": true, "Ev Idx": 13408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913728.830, "dur": 0.068,
    "args": {
      "Python parent id": 10508, "Python id": 10509, "finished": true, "Ev Idx": 13409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913729.113, "dur": 0.249,
    "args": {
      "Python parent id": 10506, "Python id": 10510, "finished": true, "Ev Idx": 13410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913729.919, "dur": 0.164,
    "args": {
      "Python parent id": 10475, "Python id": 10511, "finished": true, "Ev Idx": 13411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913730.717, "dur": 1.059,
    "args": {
      "Python parent id": 10475, "Python id": 10512, "finished": true, "Ev Idx": 13412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913731.204, "dur": 0.514,
    "args": {
      "Python parent id": 10512, "Python id": 10513, "finished": true, "Ev Idx": 13413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913732.205, "dur": 2.995,
    "args": {
      "Python parent id": 10474, "Python id": 10514, "finished": true, "Ev Idx": 13414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913736.262, "dur": 93.932,
    "args": {
      "Python parent id": 12, "Python id": 10515, "finished": true, "Ev Idx": 13415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913736.653, "dur": 0.454,
    "args": {
      "Python parent id": 10515, "Python id": 10516, "finished": true, "Ev Idx": 13416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913736.957, "dur": 0.100,
    "args": {
      "Python parent id": 10516, "Python id": 10517, "finished": true, "Ev Idx": 13417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913737.441, "dur": 0.448,
    "args": {
      "Python parent id": 10515, "Python id": 10518, "finished": true, "Ev Idx": 13418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913738.199, "dur": 0.266,
    "args": {
      "Python parent id": 10515, "Python id": 10519, "finished": true, "Ev Idx": 13419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913738.709, "dur": 1.620,
    "args": {
      "Python parent id": 10515, "Python id": 10520, "finished": true, "Ev Idx": 13420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913740.976, "dur": 89.120,
    "args": {
      "Python parent id": 10515, "Python id": 10521, "finished": true, "Ev Idx": 13421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913741.810, "dur": 88.055,
    "args": {
      "Python parent id": 10521, "Python id": 10522, "finished": true, "Ev Idx": 13422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913742.153, "dur": 87.634,
    "args": {
      "Python parent id": 10522, "Python id": 10523, "finished": true, "Ev Idx": 13423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913747.914, "dur": 80.389,
    "args": {
      "Python parent id": 10523, "Python id": 10524, "finished": true, "Ev Idx": 13424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913748.288, "dur": 79.952,
    "args": {
      "Python parent id": 10524, "Python id": 10525, "finished": true, "Ev Idx": 13425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913748.912, "dur": 79.192,
    "args": {
      "Python parent id": 10525, "Python id": 10526, "finished": true, "Ev Idx": 13426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913749.454, "dur": 0.496,
    "args": {
      "Python parent id": 10526, "Python id": 10527, "finished": true, "Ev Idx": 13427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913749.776, "dur": 0.093,
    "args": {
      "Python parent id": 10527, "Python id": 10528, "finished": true, "Ev Idx": 13428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913750.359, "dur": 0.360,
    "args": {
      "Python parent id": 10526, "Python id": 10529, "finished": true, "Ev Idx": 13429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913750.619, "dur": 0.044,
    "args": {
      "Python parent id": 10529, "Python id": 10530, "finished": true, "Ev Idx": 13430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913751.080, "dur": 76.895,
    "args": {
      "Python parent id": 10526, "Python id": 10531, "finished": true, "Ev Idx": 13431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913751.812, "dur": 74.837,
    "args": {
      "Python parent id": 10531, "Python id": 10532, "finished": true, "Ev Idx": 13432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913759.056, "dur": 19.127,
    "args": {
      "Python parent id": 10532, "Python id": 10533, "finished": true, "Ev Idx": 13433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913759.245, "dur": 0.166,
    "args": {
      "Python parent id": 10533, "Python id": 10534, "finished": true, "Ev Idx": 13434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913759.723, "dur": 18.217,
    "args": {
      "Python parent id": 10533, "Python id": 10535, "finished": true, "Ev Idx": 13435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913759.882, "dur": 8.339,
    "args": {
      "Python parent id": 10535, "Python id": 10536, "finished": true, "Ev Idx": 13436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913786.104, "dur": 6.681,
    "args": {
      "Python parent id": 10532, "Python id": 10537, "finished": true, "Ev Idx": 13437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913802.653, "dur": 8.673,
    "args": {
      "Python parent id": 10532, "Python id": 10538, "finished": true, "Ev Idx": 13438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913831.478, "dur": 18.193,
    "args": {
      "Python parent id": 12, "Python id": 10539, "finished": true, "Ev Idx": 13439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913831.866, "dur": 13.606,
    "args": {
      "Python parent id": 10539, "Python id": 10540, "finished": true, "Ev Idx": 13440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913832.081, "dur": 0.525,
    "args": {
      "Python parent id": 10540, "Python id": 10541, "finished": true, "Ev Idx": 13441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913832.258, "dur": 0.284,
    "args": {
      "Python parent id": 10541, "Python id": 10542, "finished": true, "Ev Idx": 13442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913832.400, "dur": 0.069,
    "args": {
      "Python parent id": 10542, "Python id": 10543, "finished": true, "Ev Idx": 13443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913832.886, "dur": 10.028,
    "args": {
      "Python parent id": 10540, "Python id": 10544, "finished": true, "Ev Idx": 13444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913832.970, "dur": 0.071,
    "args": {
      "Python parent id": 10544, "Python id": 10545, "finished": true, "Ev Idx": 13445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913833.180, "dur": 0.061,
    "args": {
      "Python parent id": 10544, "Python id": 10546, "finished": true, "Ev Idx": 13446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913833.403, "dur": 0.063,
    "args": {
      "Python parent id": 10544, "Python id": 10547, "finished": true, "Ev Idx": 13447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913833.923, "dur": 0.028,
    "args": {
      "Python parent id": 10544, "Python id": 10548, "finished": true, "Ev Idx": 13448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913834.267, "dur": 0.098,
    "args": {
      "Python parent id": 10544, "Python id": 10549, "finished": true, "Ev Idx": 13449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913834.727, "dur": 8.108,
    "args": {
      "Python parent id": 10544, "Python id": 10550, "finished": true, "Ev Idx": 13450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913834.843, "dur": 0.065,
    "args": {
      "Python parent id": 10550, "Python id": 10551, "finished": true, "Ev Idx": 13451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913835.057, "dur": 0.061,
    "args": {
      "Python parent id": 10550, "Python id": 10552, "finished": true, "Ev Idx": 13452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913835.299, "dur": 0.073,
    "args": {
      "Python parent id": 10550, "Python id": 10553, "finished": true, "Ev Idx": 13453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913835.742, "dur": 0.053,
    "args": {
      "Python parent id": 10550, "Python id": 10554, "finished": true, "Ev Idx": 13454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913836.097, "dur": 6.677,
    "args": {
      "Python parent id": 10550, "Python id": 10555, "finished": true, "Ev Idx": 13455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913836.328, "dur": 6.381,
    "args": {
      "Python parent id": 10555, "Python id": 10556, "finished": true, "Ev Idx": 13456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913836.499, "dur": 3.869,
    "args": {
      "Python parent id": 10556, "Python id": 10557, "finished": true, "Ev Idx": 13457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913836.718, "dur": 3.584,
    "args": {
      "Python parent id": 10557, "Python id": 10558, "finished": true, "Ev Idx": 13458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913836.875, "dur": 0.267,
    "args": {
      "Python parent id": 10558, "Python id": 10559, "finished": true, "Ev Idx": 13459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913837.017, "dur": 0.064,
    "args": {
      "Python parent id": 10559, "Python id": 10560, "finished": true, "Ev Idx": 13460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913837.324, "dur": 2.542,
    "args": {
      "Python parent id": 10558, "Python id": 10561, "finished": true, "Ev Idx": 13461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913837.541, "dur": 2.210,
    "args": {
      "Python parent id": 10561, "Python id": 10562, "finished": true, "Ev Idx": 13462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913837.777, "dur": 1.906,
    "args": {
      "Python parent id": 10562, "Python id": 10563, "finished": true, "Ev Idx": 13463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913837.977, "dur": 1.451,
    "args": {
      "Python parent id": 10563, "Python id": 10564, "finished": true, "Ev Idx": 13464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913838.205, "dur": 0.639,
    "args": {
      "Python parent id": 10564, "Python id": 10565, "finished": true, "Ev Idx": 13465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913838.310, "dur": 0.050,
    "args": {
      "Python parent id": 10565, "Python id": 10566, "finished": true, "Ev Idx": 13466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913838.608, "dur": 0.179,
    "args": {
      "Python parent id": 10565, "Python id": 10567, "finished": true, "Ev Idx": 13467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913840.105, "dur": 0.160,
    "args": {
      "Python parent id": 10558, "Python id": 10568, "finished": true, "Ev Idx": 13468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913840.929, "dur": 0.141,
    "args": {
      "Python parent id": 10556, "Python id": 10569, "finished": true, "Ev Idx": 13469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913841.354, "dur": 1.310,
    "args": {
      "Python parent id": 10556, "Python id": 10570, "finished": true, "Ev Idx": 13470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913841.519, "dur": 1.102,
    "args": {
      "Python parent id": 10570, "Python id": 10571, "finished": true, "Ev Idx": 13471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913841.683, "dur": 0.396,
    "args": {
      "Python parent id": 10571, "Python id": 10572, "finished": true, "Ev Idx": 13472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913841.827, "dur": 0.194,
    "args": {
      "Python parent id": 10572, "Python id": 10573, "finished": true, "Ev Idx": 13473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913841.909, "dur": 0.060,
    "args": {
      "Python parent id": 10573, "Python id": 10574, "finished": true, "Ev Idx": 13474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913842.253, "dur": 0.322,
    "args": {
      "Python parent id": 10571, "Python id": 10575, "finished": true, "Ev Idx": 13475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913843.069, "dur": 0.282,
    "args": {
      "Python parent id": 10540, "Python id": 10576, "finished": true, "Ev Idx": 13476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913844.026, "dur": 1.291,
    "args": {
      "Python parent id": 10540, "Python id": 10577, "finished": true, "Ev Idx": 13477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913844.595, "dur": 0.649,
    "args": {
      "Python parent id": 10577, "Python id": 10578, "finished": true, "Ev Idx": 13478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913845.874, "dur": 3.702,
    "args": {
      "Python parent id": 10539, "Python id": 10579, "finished": true, "Ev Idx": 13479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913850.299, "dur": 7.729,
    "args": {
      "Python parent id": 12, "Python id": 10580, "finished": true, "Ev Idx": 13480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740913858.559, "dur": 0.112,
    "args": {
      "Python parent id": 12, "Python id": 10581, "finished": true, "Ev Idx": 13481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913859.425, "dur": 0.235,
    "args": {
      "Python parent id": 12, "Python id": 10582, "finished": true, "Ev Idx": 13482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913860.208, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 10583, "finished": true, "Ev Idx": 13483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740913860.684, "dur": 9.428,
    "args": {
      "Python parent id": 12, "Python id": 10584, "finished": true, "Ev Idx": 13484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913871.727, "dur": 16.241,
    "args": {
      "Python parent id": 12, "Python id": 10585, "finished": true, "Ev Idx": 13485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913872.096, "dur": 11.770,
    "args": {
      "Python parent id": 10585, "Python id": 10586, "finished": true, "Ev Idx": 13486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913872.262, "dur": 0.605,
    "args": {
      "Python parent id": 10586, "Python id": 10587, "finished": true, "Ev Idx": 13487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913872.452, "dur": 0.320,
    "args": {
      "Python parent id": 10587, "Python id": 10588, "finished": true, "Ev Idx": 13488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913872.582, "dur": 0.122,
    "args": {
      "Python parent id": 10588, "Python id": 10589, "finished": true, "Ev Idx": 13489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913873.151, "dur": 8.694,
    "args": {
      "Python parent id": 10586, "Python id": 10590, "finished": true, "Ev Idx": 13490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913873.274, "dur": 0.066,
    "args": {
      "Python parent id": 10590, "Python id": 10591, "finished": true, "Ev Idx": 13491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913873.458, "dur": 0.087,
    "args": {
      "Python parent id": 10590, "Python id": 10592, "finished": true, "Ev Idx": 13492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913873.716, "dur": 0.059,
    "args": {
      "Python parent id": 10590, "Python id": 10593, "finished": true, "Ev Idx": 13493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913874.194, "dur": 0.041,
    "args": {
      "Python parent id": 10590, "Python id": 10594, "finished": true, "Ev Idx": 13494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913874.514, "dur": 0.084,
    "args": {
      "Python parent id": 10590, "Python id": 10595, "finished": true, "Ev Idx": 13495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913874.990, "dur": 6.798,
    "args": {
      "Python parent id": 10590, "Python id": 10596, "finished": true, "Ev Idx": 13496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913875.114, "dur": 0.065,
    "args": {
      "Python parent id": 10596, "Python id": 10597, "finished": true, "Ev Idx": 13497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913875.332, "dur": 0.068,
    "args": {
      "Python parent id": 10596, "Python id": 10598, "finished": true, "Ev Idx": 13498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913875.595, "dur": 0.081,
    "args": {
      "Python parent id": 10596, "Python id": 10599, "finished": true, "Ev Idx": 13499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913876.004, "dur": 0.042,
    "args": {
      "Python parent id": 10596, "Python id": 10600, "finished": true, "Ev Idx": 13500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913876.350, "dur": 5.326,
    "args": {
      "Python parent id": 10596, "Python id": 10601, "finished": true, "Ev Idx": 13501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913876.561, "dur": 5.056,
    "args": {
      "Python parent id": 10601, "Python id": 10602, "finished": true, "Ev Idx": 13502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913876.720, "dur": 3.247,
    "args": {
      "Python parent id": 10602, "Python id": 10603, "finished": true, "Ev Idx": 13503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913876.963, "dur": 2.959,
    "args": {
      "Python parent id": 10603, "Python id": 10604, "finished": true, "Ev Idx": 13504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913877.101, "dur": 0.267,
    "args": {
      "Python parent id": 10604, "Python id": 10605, "finished": true, "Ev Idx": 13505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913877.248, "dur": 0.069,
    "args": {
      "Python parent id": 10605, "Python id": 10606, "finished": true, "Ev Idx": 13506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913877.553, "dur": 2.070,
    "args": {
      "Python parent id": 10604, "Python id": 10607, "finished": true, "Ev Idx": 13507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913877.776, "dur": 1.763,
    "args": {
      "Python parent id": 10607, "Python id": 10608, "finished": true, "Ev Idx": 13508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913877.984, "dur": 1.507,
    "args": {
      "Python parent id": 10608, "Python id": 10609, "finished": true, "Ev Idx": 13509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913878.112, "dur": 1.119,
    "args": {
      "Python parent id": 10609, "Python id": 10610, "finished": true, "Ev Idx": 13510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913878.339, "dur": 0.487,
    "args": {
      "Python parent id": 10610, "Python id": 10611, "finished": true, "Ev Idx": 13511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913878.427, "dur": 0.053,
    "args": {
      "Python parent id": 10611, "Python id": 10612, "finished": true, "Ev Idx": 13512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913878.639, "dur": 0.142,
    "args": {
      "Python parent id": 10611, "Python id": 10613, "finished": true, "Ev Idx": 13513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913879.815, "dur": 0.072,
    "args": {
      "Python parent id": 10604, "Python id": 10614, "finished": true, "Ev Idx": 13514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.135, "dur": 0.075,
    "args": {
      "Python parent id": 10602, "Python id": 10615, "finished": true, "Ev Idx": 13515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.450, "dur": 1.124,
    "args": {
      "Python parent id": 10602, "Python id": 10616, "finished": true, "Ev Idx": 13516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.586, "dur": 0.944,
    "args": {
      "Python parent id": 10616, "Python id": 10617, "finished": true, "Ev Idx": 13517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.724, "dur": 0.391,
    "args": {
      "Python parent id": 10617, "Python id": 10618, "finished": true, "Ev Idx": 13518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.863, "dur": 0.195,
    "args": {
      "Python parent id": 10618, "Python id": 10619, "finished": true, "Ev Idx": 13519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913880.968, "dur": 0.058,
    "args": {
      "Python parent id": 10619, "Python id": 10620, "finished": true, "Ev Idx": 13520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913881.246, "dur": 0.242,
    "args": {
      "Python parent id": 10617, "Python id": 10621, "finished": true, "Ev Idx": 13521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913881.988, "dur": 0.155,
    "args": {
      "Python parent id": 10586, "Python id": 10622, "finished": true, "Ev Idx": 13522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913882.743, "dur": 0.959,
    "args": {
      "Python parent id": 10586, "Python id": 10623, "finished": true, "Ev Idx": 13523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913883.153, "dur": 0.478,
    "args": {
      "Python parent id": 10623, "Python id": 10624, "finished": true, "Ev Idx": 13524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913884.123, "dur": 3.784,
    "args": {
      "Python parent id": 10585, "Python id": 10625, "finished": true, "Ev Idx": 13525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740913888.948, "dur": 90.797,
    "args": {
      "Python parent id": 12, "Python id": 10626, "finished": true, "Ev Idx": 13526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913889.359, "dur": 0.443,
    "args": {
      "Python parent id": 10626, "Python id": 10627, "finished": true, "Ev Idx": 13527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913889.657, "dur": 0.086,
    "args": {
      "Python parent id": 10627, "Python id": 10628, "finished": true, "Ev Idx": 13528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913890.160, "dur": 0.501,
    "args": {
      "Python parent id": 10626, "Python id": 10629, "finished": true, "Ev Idx": 13529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740913890.912, "dur": 0.325,
    "args": {
      "Python parent id": 10626, "Python id": 10630, "finished": true, "Ev Idx": 13530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740913891.462, "dur": 1.522,
    "args": {
      "Python parent id": 10626, "Python id": 10631, "finished": true, "Ev Idx": 13531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740913893.606, "dur": 86.031,
    "args": {
      "Python parent id": 10626, "Python id": 10632, "finished": true, "Ev Idx": 13532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740913894.458, "dur": 84.939,
    "args": {
      "Python parent id": 10632, "Python id": 10633, "finished": true, "Ev Idx": 13533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913894.837, "dur": 84.499,
    "args": {
      "Python parent id": 10633, "Python id": 10634, "finished": true, "Ev Idx": 13534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740913900.136, "dur": 77.768,
    "args": {
      "Python parent id": 10634, "Python id": 10635, "finished": true, "Ev Idx": 13535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740913900.481, "dur": 77.320,
    "args": {
      "Python parent id": 10635, "Python id": 10636, "finished": true, "Ev Idx": 13536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740913901.033, "dur": 76.568,
    "args": {
      "Python parent id": 10636, "Python id": 10637, "finished": true, "Ev Idx": 13537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913901.522, "dur": 0.489,
    "args": {
      "Python parent id": 10637, "Python id": 10638, "finished": true, "Ev Idx": 13538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913901.835, "dur": 0.047,
    "args": {
      "Python parent id": 10638, "Python id": 10639, "finished": true, "Ev Idx": 13539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740913902.308, "dur": 0.334,
    "args": {
      "Python parent id": 10637, "Python id": 10640, "finished": true, "Ev Idx": 13540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740913902.523, "dur": 0.075,
    "args": {
      "Python parent id": 10640, "Python id": 10641, "finished": true, "Ev Idx": 13541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913903.041, "dur": 74.492,
    "args": {
      "Python parent id": 10637, "Python id": 10642, "finished": true, "Ev Idx": 13542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740913903.771, "dur": 72.302,
    "args": {
      "Python parent id": 10642, "Python id": 10643, "finished": true, "Ev Idx": 13543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740913910.912, "dur": 17.858,
    "args": {
      "Python parent id": 10643, "Python id": 10644, "finished": true, "Ev Idx": 13544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740913911.098, "dur": 0.162,
    "args": {
      "Python parent id": 10644, "Python id": 10645, "finished": true, "Ev Idx": 13545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740913911.632, "dur": 17.020,
    "args": {
      "Python parent id": 10644, "Python id": 10646, "finished": true, "Ev Idx": 13546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913911.748, "dur": 7.819,
    "args": {
      "Python parent id": 10646, "Python id": 10647, "finished": true, "Ev Idx": 13547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913936.527, "dur": 7.316,
    "args": {
      "Python parent id": 10643, "Python id": 10648, "finished": true, "Ev Idx": 13548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740913952.321, "dur": 8.294,
    "args": {
      "Python parent id": 10643, "Python id": 10649, "finished": true, "Ev Idx": 13549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740913980.926, "dur": 17.410,
    "args": {
      "Python parent id": 12, "Python id": 10650, "finished": true, "Ev Idx": 13550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740913981.298, "dur": 13.013,
    "args": {
      "Python parent id": 10650, "Python id": 10651, "finished": true, "Ev Idx": 13551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913981.523, "dur": 0.530,
    "args": {
      "Python parent id": 10651, "Python id": 10652, "finished": true, "Ev Idx": 13552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913981.711, "dur": 0.276,
    "args": {
      "Python parent id": 10652, "Python id": 10653, "finished": true, "Ev Idx": 13553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913981.841, "dur": 0.086,
    "args": {
      "Python parent id": 10653, "Python id": 10654, "finished": true, "Ev Idx": 13554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913982.412, "dur": 9.348,
    "args": {
      "Python parent id": 10651, "Python id": 10655, "finished": true, "Ev Idx": 13555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913982.537, "dur": 0.100,
    "args": {
      "Python parent id": 10655, "Python id": 10656, "finished": true, "Ev Idx": 13556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913982.784, "dur": 0.063,
    "args": {
      "Python parent id": 10655, "Python id": 10657, "finished": true, "Ev Idx": 13557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913983.007, "dur": 0.062,
    "args": {
      "Python parent id": 10655, "Python id": 10658, "finished": true, "Ev Idx": 13558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913983.533, "dur": 0.047,
    "args": {
      "Python parent id": 10655, "Python id": 10659, "finished": true, "Ev Idx": 13559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913983.869, "dur": 0.084,
    "args": {
      "Python parent id": 10655, "Python id": 10660, "finished": true, "Ev Idx": 13560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913984.322, "dur": 7.381,
    "args": {
      "Python parent id": 10655, "Python id": 10661, "finished": true, "Ev Idx": 13561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913984.461, "dur": 0.067,
    "args": {
      "Python parent id": 10661, "Python id": 10662, "finished": true, "Ev Idx": 13562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913984.720, "dur": 0.060,
    "args": {
      "Python parent id": 10661, "Python id": 10663, "finished": true, "Ev Idx": 13563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913984.989, "dur": 0.076,
    "args": {
      "Python parent id": 10661, "Python id": 10664, "finished": true, "Ev Idx": 13564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740913985.396, "dur": 0.040,
    "args": {
      "Python parent id": 10661, "Python id": 10665, "finished": true, "Ev Idx": 13565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740913985.770, "dur": 5.841,
    "args": {
      "Python parent id": 10661, "Python id": 10666, "finished": true, "Ev Idx": 13566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740913985.974, "dur": 5.574,
    "args": {
      "Python parent id": 10666, "Python id": 10667, "finished": true, "Ev Idx": 13567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740913986.116, "dur": 3.523,
    "args": {
      "Python parent id": 10667, "Python id": 10668, "finished": true, "Ev Idx": 13568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740913986.308, "dur": 3.288,
    "args": {
      "Python parent id": 10668, "Python id": 10669, "finished": true, "Ev Idx": 13569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740913986.471, "dur": 0.258,
    "args": {
      "Python parent id": 10669, "Python id": 10670, "finished": true, "Ev Idx": 13570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740913986.603, "dur": 0.074,
    "args": {
      "Python parent id": 10670, "Python id": 10671, "finished": true, "Ev Idx": 13571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740913986.887, "dur": 2.300,
    "args": {
      "Python parent id": 10669, "Python id": 10672, "finished": true, "Ev Idx": 13572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740913987.153, "dur": 1.915,
    "args": {
      "Python parent id": 10672, "Python id": 10673, "finished": true, "Ev Idx": 13573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740913987.393, "dur": 1.615,
    "args": {
      "Python parent id": 10673, "Python id": 10674, "finished": true, "Ev Idx": 13574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740913987.539, "dur": 1.267,
    "args": {
      "Python parent id": 10674, "Python id": 10675, "finished": true, "Ev Idx": 13575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740913987.752, "dur": 0.552,
    "args": {
      "Python parent id": 10675, "Python id": 10676, "finished": true, "Ev Idx": 13576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740913987.860, "dur": 0.043,
    "args": {
      "Python parent id": 10676, "Python id": 10677, "finished": true, "Ev Idx": 13577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740913988.082, "dur": 0.167,
    "args": {
      "Python parent id": 10676, "Python id": 10678, "finished": true, "Ev Idx": 13578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740913989.393, "dur": 0.158,
    "args": {
      "Python parent id": 10669, "Python id": 10679, "finished": true, "Ev Idx": 13579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740913989.802, "dur": 0.143,
    "args": {
      "Python parent id": 10667, "Python id": 10680, "finished": true, "Ev Idx": 13580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740913990.265, "dur": 1.237,
    "args": {
      "Python parent id": 10667, "Python id": 10681, "finished": true, "Ev Idx": 13581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740913990.419, "dur": 1.039,
    "args": {
      "Python parent id": 10681, "Python id": 10682, "finished": true, "Ev Idx": 13582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740913990.561, "dur": 0.404,
    "args": {
      "Python parent id": 10682, "Python id": 10683, "finished": true, "Ev Idx": 13583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740913990.699, "dur": 0.208,
    "args": {
      "Python parent id": 10683, "Python id": 10684, "finished": true, "Ev Idx": 13584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740913990.801, "dur": 0.052,
    "args": {
      "Python parent id": 10684, "Python id": 10685, "finished": true, "Ev Idx": 13585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740913991.097, "dur": 0.320,
    "args": {
      "Python parent id": 10682, "Python id": 10686, "finished": true, "Ev Idx": 13586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740913991.911, "dur": 0.269,
    "args": {
      "Python parent id": 10651, "Python id": 10687, "finished": true, "Ev Idx": 13587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740913992.869, "dur": 1.286,
    "args": {
      "Python parent id": 10651, "Python id": 10688, "finished": true, "Ev Idx": 13588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740913993.462, "dur": 0.622,
    "args": {
      "Python parent id": 10688, "Python id": 10689, "finished": true, "Ev Idx": 13589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740913994.614, "dur": 3.666,
    "args": {
      "Python parent id": 10650, "Python id": 10690, "finished": true, "Ev Idx": 13590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740913998.937, "dur": 7.389,
    "args": {
      "Python parent id": 12, "Python id": 10691, "finished": true, "Ev Idx": 13591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914006.800, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 10692, "finished": true, "Ev Idx": 13592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914007.760, "dur": 0.166,
    "args": {
      "Python parent id": 12, "Python id": 10693, "finished": true, "Ev Idx": 13593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914008.446, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 10694, "finished": true, "Ev Idx": 13594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914008.909, "dur": 11.291,
    "args": {
      "Python parent id": 12, "Python id": 10695, "finished": true, "Ev Idx": 13595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914021.499, "dur": 15.717,
    "args": {
      "Python parent id": 12, "Python id": 10696, "finished": true, "Ev Idx": 13596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914021.837, "dur": 11.359,
    "args": {
      "Python parent id": 10696, "Python id": 10697, "finished": true, "Ev Idx": 13597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914021.980, "dur": 0.508,
    "args": {
      "Python parent id": 10697, "Python id": 10698, "finished": true, "Ev Idx": 13598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914022.128, "dur": 0.282,
    "args": {
      "Python parent id": 10698, "Python id": 10699, "finished": true, "Ev Idx": 13599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914022.255, "dur": 0.097,
    "args": {
      "Python parent id": 10699, "Python id": 10700, "finished": true, "Ev Idx": 13600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914022.738, "dur": 8.571,
    "args": {
      "Python parent id": 10697, "Python id": 10701, "finished": true, "Ev Idx": 13601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914022.826, "dur": 0.068,
    "args": {
      "Python parent id": 10701, "Python id": 10702, "finished": true, "Ev Idx": 13602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914023.011, "dur": 0.066,
    "args": {
      "Python parent id": 10701, "Python id": 10703, "finished": true, "Ev Idx": 13603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914023.254, "dur": 0.066,
    "args": {
      "Python parent id": 10701, "Python id": 10704, "finished": true, "Ev Idx": 13604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914023.736, "dur": 0.045,
    "args": {
      "Python parent id": 10701, "Python id": 10705, "finished": true, "Ev Idx": 13605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914024.084, "dur": 0.083,
    "args": {
      "Python parent id": 10701, "Python id": 10706, "finished": true, "Ev Idx": 13606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914024.579, "dur": 6.659,
    "args": {
      "Python parent id": 10701, "Python id": 10707, "finished": true, "Ev Idx": 13607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914024.669, "dur": 0.065,
    "args": {
      "Python parent id": 10707, "Python id": 10708, "finished": true, "Ev Idx": 13608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914024.882, "dur": 0.068,
    "args": {
      "Python parent id": 10707, "Python id": 10709, "finished": true, "Ev Idx": 13609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914025.159, "dur": 0.067,
    "args": {
      "Python parent id": 10707, "Python id": 10710, "finished": true, "Ev Idx": 13610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914025.548, "dur": 0.069,
    "args": {
      "Python parent id": 10707, "Python id": 10711, "finished": true, "Ev Idx": 13611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914025.912, "dur": 5.268,
    "args": {
      "Python parent id": 10707, "Python id": 10712, "finished": true, "Ev Idx": 13612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914026.189, "dur": 4.929,
    "args": {
      "Python parent id": 10712, "Python id": 10713, "finished": true, "Ev Idx": 13613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914026.312, "dur": 3.157,
    "args": {
      "Python parent id": 10713, "Python id": 10714, "finished": true, "Ev Idx": 13614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914026.500, "dur": 2.923,
    "args": {
      "Python parent id": 10714, "Python id": 10715, "finished": true, "Ev Idx": 13615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914026.640, "dur": 0.276,
    "args": {
      "Python parent id": 10715, "Python id": 10716, "finished": true, "Ev Idx": 13616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914026.772, "dur": 0.062,
    "args": {
      "Python parent id": 10716, "Python id": 10717, "finished": true, "Ev Idx": 13617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.102, "dur": 2.026,
    "args": {
      "Python parent id": 10715, "Python id": 10718, "finished": true, "Ev Idx": 13618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.299, "dur": 1.766,
    "args": {
      "Python parent id": 10718, "Python id": 10719, "finished": true, "Ev Idx": 13619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.488, "dur": 1.530,
    "args": {
      "Python parent id": 10719, "Python id": 10720, "finished": true, "Ev Idx": 13620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.672, "dur": 1.145,
    "args": {
      "Python parent id": 10720, "Python id": 10721, "finished": true, "Ev Idx": 13621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.841, "dur": 0.528,
    "args": {
      "Python parent id": 10721, "Python id": 10722, "finished": true, "Ev Idx": 13622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914027.922, "dur": 0.055,
    "args": {
      "Python parent id": 10722, "Python id": 10723, "finished": true, "Ev Idx": 13623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914028.168, "dur": 0.146,
    "args": {
      "Python parent id": 10722, "Python id": 10724, "finished": true, "Ev Idx": 13624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914029.317, "dur": 0.072,
    "args": {
      "Python parent id": 10715, "Python id": 10725, "finished": true, "Ev Idx": 13625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914029.625, "dur": 0.079,
    "args": {
      "Python parent id": 10713, "Python id": 10726, "finished": true, "Ev Idx": 13626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914029.972, "dur": 1.102,
    "args": {
      "Python parent id": 10713, "Python id": 10727, "finished": true, "Ev Idx": 13627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914030.121, "dur": 0.908,
    "args": {
      "Python parent id": 10727, "Python id": 10728, "finished": true, "Ev Idx": 13628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914030.260, "dur": 0.376,
    "args": {
      "Python parent id": 10728, "Python id": 10729, "finished": true, "Ev Idx": 13629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914030.394, "dur": 0.182,
    "args": {
      "Python parent id": 10729, "Python id": 10730, "finished": true, "Ev Idx": 13630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914030.488, "dur": 0.057,
    "args": {
      "Python parent id": 10730, "Python id": 10731, "finished": true, "Ev Idx": 13631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914030.765, "dur": 0.235,
    "args": {
      "Python parent id": 10728, "Python id": 10732, "finished": true, "Ev Idx": 13632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914031.426, "dur": 0.149,
    "args": {
      "Python parent id": 10697, "Python id": 10733, "finished": true, "Ev Idx": 13633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914032.194, "dur": 0.870,
    "args": {
      "Python parent id": 10697, "Python id": 10734, "finished": true, "Ev Idx": 13634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914032.584, "dur": 0.431,
    "args": {
      "Python parent id": 10734, "Python id": 10735, "finished": true, "Ev Idx": 13635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914033.440, "dur": 3.730,
    "args": {
      "Python parent id": 10696, "Python id": 10736, "finished": true, "Ev Idx": 13636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914038.164, "dur": 138.801,
    "args": {
      "Python parent id": 12, "Python id": 10737, "finished": true, "Ev Idx": 13637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914038.579, "dur": 0.445,
    "args": {
      "Python parent id": 10737, "Python id": 10738, "finished": true, "Ev Idx": 13638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914038.858, "dur": 0.084,
    "args": {
      "Python parent id": 10738, "Python id": 10739, "finished": true, "Ev Idx": 13639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914039.328, "dur": 0.432,
    "args": {
      "Python parent id": 10737, "Python id": 10740, "finished": true, "Ev Idx": 13640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914039.999, "dur": 0.271,
    "args": {
      "Python parent id": 10737, "Python id": 10741, "finished": true, "Ev Idx": 13641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914040.498, "dur": 1.540,
    "args": {
      "Python parent id": 10737, "Python id": 10742, "finished": true, "Ev Idx": 13642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914042.585, "dur": 134.275,
    "args": {
      "Python parent id": 10737, "Python id": 10743, "finished": true, "Ev Idx": 13643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914043.361, "dur": 133.274,
    "args": {
      "Python parent id": 10743, "Python id": 10744, "finished": true, "Ev Idx": 13644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914043.674, "dur": 132.880,
    "args": {
      "Python parent id": 10744, "Python id": 10745, "finished": true, "Ev Idx": 13645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914048.708, "dur": 126.052,
    "args": {
      "Python parent id": 10745, "Python id": 10746, "finished": true, "Ev Idx": 13646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914049.046, "dur": 125.626,
    "args": {
      "Python parent id": 10746, "Python id": 10747, "finished": true, "Ev Idx": 13647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914049.626, "dur": 124.914,
    "args": {
      "Python parent id": 10747, "Python id": 10748, "finished": true, "Ev Idx": 13648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914050.159, "dur": 0.408,
    "args": {
      "Python parent id": 10748, "Python id": 10749, "finished": true, "Ev Idx": 13649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914050.422, "dur": 0.083,
    "args": {
      "Python parent id": 10749, "Python id": 10750, "finished": true, "Ev Idx": 13650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914050.844, "dur": 0.333,
    "args": {
      "Python parent id": 10748, "Python id": 10751, "finished": true, "Ev Idx": 13651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914051.080, "dur": 0.040,
    "args": {
      "Python parent id": 10751, "Python id": 10752, "finished": true, "Ev Idx": 13652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914051.514, "dur": 122.800,
    "args": {
      "Python parent id": 10748, "Python id": 10753, "finished": true, "Ev Idx": 13653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914052.206, "dur": 120.390,
    "args": {
      "Python parent id": 10753, "Python id": 10754, "finished": true, "Ev Idx": 13654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914059.241, "dur": 56.752,
    "args": {
      "Python parent id": 10754, "Python id": 10755, "finished": true, "Ev Idx": 13655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914059.404, "dur": 0.158,
    "args": {
      "Python parent id": 10755, "Python id": 10756, "finished": true, "Ev Idx": 13656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914059.910, "dur": 55.840,
    "args": {
      "Python parent id": 10755, "Python id": 10757, "finished": true, "Ev Idx": 13657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914060.012, "dur": 7.986,
    "args": {
      "Python parent id": 10757, "Python id": 10758, "finished": true, "Ev Idx": 13658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914124.941, "dur": 9.336,
    "args": {
      "Python parent id": 10754, "Python id": 10759, "finished": true, "Ev Idx": 13659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914145.111, "dur": 10.814,
    "args": {
      "Python parent id": 10754, "Python id": 10760, "finished": true, "Ev Idx": 13660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914178.643, "dur": 22.144,
    "args": {
      "Python parent id": 12, "Python id": 10761, "finished": true, "Ev Idx": 13661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914178.997, "dur": 16.823,
    "args": {
      "Python parent id": 10761, "Python id": 10762, "finished": true, "Ev Idx": 13662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914179.219, "dur": 0.801,
    "args": {
      "Python parent id": 10762, "Python id": 10763, "finished": true, "Ev Idx": 13663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914179.481, "dur": 0.436,
    "args": {
      "Python parent id": 10763, "Python id": 10764, "finished": true, "Ev Idx": 13664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914179.699, "dur": 0.148,
    "args": {
      "Python parent id": 10764, "Python id": 10765, "finished": true, "Ev Idx": 13665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914180.443, "dur": 12.031,
    "args": {
      "Python parent id": 10762, "Python id": 10766, "finished": true, "Ev Idx": 13666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914180.558, "dur": 0.107,
    "args": {
      "Python parent id": 10766, "Python id": 10767, "finished": true, "Ev Idx": 13667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914180.812, "dur": 0.067,
    "args": {
      "Python parent id": 10766, "Python id": 10768, "finished": true, "Ev Idx": 13668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914181.084, "dur": 0.059,
    "args": {
      "Python parent id": 10766, "Python id": 10769, "finished": true, "Ev Idx": 13669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914181.663, "dur": 0.344,
    "args": {
      "Python parent id": 10766, "Python id": 10770, "finished": true, "Ev Idx": 13670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914182.430, "dur": 0.108,
    "args": {
      "Python parent id": 10766, "Python id": 10771, "finished": true, "Ev Idx": 13671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914182.999, "dur": 9.417,
    "args": {
      "Python parent id": 10766, "Python id": 10772, "finished": true, "Ev Idx": 13672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914183.127, "dur": 0.064,
    "args": {
      "Python parent id": 10772, "Python id": 10773, "finished": true, "Ev Idx": 13673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914183.420, "dur": 0.059,
    "args": {
      "Python parent id": 10772, "Python id": 10774, "finished": true, "Ev Idx": 13674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914183.674, "dur": 0.078,
    "args": {
      "Python parent id": 10772, "Python id": 10775, "finished": true, "Ev Idx": 13675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914184.068, "dur": 0.074,
    "args": {
      "Python parent id": 10772, "Python id": 10776, "finished": true, "Ev Idx": 13676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914184.506, "dur": 7.788,
    "args": {
      "Python parent id": 10772, "Python id": 10777, "finished": true, "Ev Idx": 13677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914184.886, "dur": 7.322,
    "args": {
      "Python parent id": 10777, "Python id": 10778, "finished": true, "Ev Idx": 13678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914185.055, "dur": 4.783,
    "args": {
      "Python parent id": 10778, "Python id": 10779, "finished": true, "Ev Idx": 13679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914185.345, "dur": 4.404,
    "args": {
      "Python parent id": 10779, "Python id": 10780, "finished": true, "Ev Idx": 13680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914185.483, "dur": 0.356,
    "args": {
      "Python parent id": 10780, "Python id": 10781, "finished": true, "Ev Idx": 13681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914185.664, "dur": 0.081,
    "args": {
      "Python parent id": 10781, "Python id": 10782, "finished": true, "Ev Idx": 13682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914186.034, "dur": 3.237,
    "args": {
      "Python parent id": 10780, "Python id": 10783, "finished": true, "Ev Idx": 13683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914186.311, "dur": 2.811,
    "args": {
      "Python parent id": 10783, "Python id": 10784, "finished": true, "Ev Idx": 13684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914186.503, "dur": 2.552,
    "args": {
      "Python parent id": 10784, "Python id": 10785, "finished": true, "Ev Idx": 13685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914186.735, "dur": 1.913,
    "args": {
      "Python parent id": 10785, "Python id": 10786, "finished": true, "Ev Idx": 13686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914187.019, "dur": 0.755,
    "args": {
      "Python parent id": 10786, "Python id": 10787, "finished": true, "Ev Idx": 13687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914187.134, "dur": 0.098,
    "args": {
      "Python parent id": 10787, "Python id": 10788, "finished": true, "Ev Idx": 13688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914187.459, "dur": 0.277,
    "args": {
      "Python parent id": 10787, "Python id": 10789, "finished": true, "Ev Idx": 13689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914189.507, "dur": 0.199,
    "args": {
      "Python parent id": 10780, "Python id": 10790, "finished": true, "Ev Idx": 13690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914190.102, "dur": 0.149,
    "args": {
      "Python parent id": 10778, "Python id": 10791, "finished": true, "Ev Idx": 13691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914190.588, "dur": 1.553,
    "args": {
      "Python parent id": 10778, "Python id": 10792, "finished": true, "Ev Idx": 13692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914190.823, "dur": 1.258,
    "args": {
      "Python parent id": 10792, "Python id": 10793, "finished": true, "Ev Idx": 13693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914190.965, "dur": 0.505,
    "args": {
      "Python parent id": 10793, "Python id": 10794, "finished": true, "Ev Idx": 13694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914191.121, "dur": 0.261,
    "args": {
      "Python parent id": 10794, "Python id": 10795, "finished": true, "Ev Idx": 13695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914191.250, "dur": 0.079,
    "args": {
      "Python parent id": 10795, "Python id": 10796, "finished": true, "Ev Idx": 13696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914191.624, "dur": 0.411,
    "args": {
      "Python parent id": 10793, "Python id": 10797, "finished": true, "Ev Idx": 13697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914192.665, "dur": 0.351,
    "args": {
      "Python parent id": 10762, "Python id": 10798, "finished": true, "Ev Idx": 13698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914193.922, "dur": 1.688,
    "args": {
      "Python parent id": 10762, "Python id": 10799, "finished": true, "Ev Idx": 13699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914194.705, "dur": 0.814,
    "args": {
      "Python parent id": 10799, "Python id": 10800, "finished": true, "Ev Idx": 13700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914196.285, "dur": 4.385,
    "args": {
      "Python parent id": 10761, "Python id": 10801, "finished": true, "Ev Idx": 13701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914201.444, "dur": 8.788,
    "args": {
      "Python parent id": 12, "Python id": 10802, "finished": true, "Ev Idx": 13702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914210.730, "dur": 0.138,
    "args": {
      "Python parent id": 12, "Python id": 10803, "finished": true, "Ev Idx": 13703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914211.763, "dur": 0.235,
    "args": {
      "Python parent id": 12, "Python id": 10804, "finished": true, "Ev Idx": 13704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914212.505, "dur": 0.082,
    "args": {
      "Python parent id": 12, "Python id": 10805, "finished": true, "Ev Idx": 13705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914212.978, "dur": 12.175,
    "args": {
      "Python parent id": 12, "Python id": 10806, "finished": true, "Ev Idx": 13706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914226.569, "dur": 16.699,
    "args": {
      "Python parent id": 12, "Python id": 10807, "finished": true, "Ev Idx": 13707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914226.921, "dur": 12.444,
    "args": {
      "Python parent id": 10807, "Python id": 10808, "finished": true, "Ev Idx": 13708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914227.084, "dur": 0.634,
    "args": {
      "Python parent id": 10808, "Python id": 10809, "finished": true, "Ev Idx": 13709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914227.259, "dur": 0.359,
    "args": {
      "Python parent id": 10809, "Python id": 10810, "finished": true, "Ev Idx": 13710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914227.400, "dur": 0.162,
    "args": {
      "Python parent id": 10810, "Python id": 10811, "finished": true, "Ev Idx": 13711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914228.097, "dur": 9.121,
    "args": {
      "Python parent id": 10808, "Python id": 10812, "finished": true, "Ev Idx": 13712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914228.178, "dur": 0.068,
    "args": {
      "Python parent id": 10812, "Python id": 10813, "finished": true, "Ev Idx": 13713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914228.345, "dur": 0.058,
    "args": {
      "Python parent id": 10812, "Python id": 10814, "finished": true, "Ev Idx": 13714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914228.550, "dur": 0.058,
    "args": {
      "Python parent id": 10812, "Python id": 10815, "finished": true, "Ev Idx": 13715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914228.950, "dur": 0.075,
    "args": {
      "Python parent id": 10812, "Python id": 10816, "finished": true, "Ev Idx": 13716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914229.332, "dur": 0.072,
    "args": {
      "Python parent id": 10812, "Python id": 10817, "finished": true, "Ev Idx": 13717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914229.775, "dur": 7.380,
    "args": {
      "Python parent id": 10812, "Python id": 10818, "finished": true, "Ev Idx": 13718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914229.867, "dur": 0.089,
    "args": {
      "Python parent id": 10818, "Python id": 10819, "finished": true, "Ev Idx": 13719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914230.116, "dur": 0.067,
    "args": {
      "Python parent id": 10818, "Python id": 10820, "finished": true, "Ev Idx": 13720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914230.412, "dur": 0.072,
    "args": {
      "Python parent id": 10818, "Python id": 10821, "finished": true, "Ev Idx": 13721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914230.848, "dur": 0.052,
    "args": {
      "Python parent id": 10818, "Python id": 10822, "finished": true, "Ev Idx": 13722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914231.220, "dur": 5.878,
    "args": {
      "Python parent id": 10818, "Python id": 10823, "finished": true, "Ev Idx": 13723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914231.457, "dur": 5.580,
    "args": {
      "Python parent id": 10823, "Python id": 10824, "finished": true, "Ev Idx": 13724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914231.583, "dur": 3.566,
    "args": {
      "Python parent id": 10824, "Python id": 10825, "finished": true, "Ev Idx": 13725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914231.766, "dur": 3.312,
    "args": {
      "Python parent id": 10825, "Python id": 10826, "finished": true, "Ev Idx": 13726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914231.900, "dur": 0.309,
    "args": {
      "Python parent id": 10826, "Python id": 10827, "finished": true, "Ev Idx": 13727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914232.068, "dur": 0.075,
    "args": {
      "Python parent id": 10827, "Python id": 10828, "finished": true, "Ev Idx": 13728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914232.398, "dur": 2.393,
    "args": {
      "Python parent id": 10826, "Python id": 10829, "finished": true, "Ev Idx": 13729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914232.599, "dur": 2.088,
    "args": {
      "Python parent id": 10829, "Python id": 10830, "finished": true, "Ev Idx": 13730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914232.827, "dur": 1.811,
    "args": {
      "Python parent id": 10830, "Python id": 10831, "finished": true, "Ev Idx": 13731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914232.986, "dur": 1.364,
    "args": {
      "Python parent id": 10831, "Python id": 10832, "finished": true, "Ev Idx": 13732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914233.223, "dur": 0.582,
    "args": {
      "Python parent id": 10832, "Python id": 10833, "finished": true, "Ev Idx": 13733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914233.326, "dur": 0.068,
    "args": {
      "Python parent id": 10833, "Python id": 10834, "finished": true, "Ev Idx": 13734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914233.586, "dur": 0.181,
    "args": {
      "Python parent id": 10833, "Python id": 10835, "finished": true, "Ev Idx": 13735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914234.967, "dur": 0.072,
    "args": {
      "Python parent id": 10826, "Python id": 10836, "finished": true, "Ev Idx": 13736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914235.358, "dur": 0.113,
    "args": {
      "Python parent id": 10824, "Python id": 10837, "finished": true, "Ev Idx": 13737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914235.781, "dur": 1.211,
    "args": {
      "Python parent id": 10824, "Python id": 10838, "finished": true, "Ev Idx": 13738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914235.958, "dur": 0.981,
    "args": {
      "Python parent id": 10838, "Python id": 10839, "finished": true, "Ev Idx": 13739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914236.092, "dur": 0.412,
    "args": {
      "Python parent id": 10839, "Python id": 10840, "finished": true, "Ev Idx": 13740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914236.267, "dur": 0.177,
    "args": {
      "Python parent id": 10840, "Python id": 10841, "finished": true, "Ev Idx": 13741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914236.355, "dur": 0.057,
    "args": {
      "Python parent id": 10841, "Python id": 10842, "finished": true, "Ev Idx": 13742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914236.630, "dur": 0.252,
    "args": {
      "Python parent id": 10839, "Python id": 10843, "finished": true, "Ev Idx": 13743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914237.341, "dur": 0.205,
    "args": {
      "Python parent id": 10808, "Python id": 10844, "finished": true, "Ev Idx": 13744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914238.202, "dur": 1.034,
    "args": {
      "Python parent id": 10808, "Python id": 10845, "finished": true, "Ev Idx": 13745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914238.676, "dur": 0.491,
    "args": {
      "Python parent id": 10845, "Python id": 10846, "finished": true, "Ev Idx": 13746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914239.662, "dur": 3.546,
    "args": {
      "Python parent id": 10807, "Python id": 10847, "finished": true, "Ev Idx": 13747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914244.248, "dur": 101.128,
    "args": {
      "Python parent id": 12, "Python id": 10848, "finished": true, "Ev Idx": 13748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914244.695, "dur": 0.499,
    "args": {
      "Python parent id": 10848, "Python id": 10849, "finished": true, "Ev Idx": 13749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914245.037, "dur": 0.094,
    "args": {
      "Python parent id": 10849, "Python id": 10850, "finished": true, "Ev Idx": 13750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914245.506, "dur": 0.488,
    "args": {
      "Python parent id": 10848, "Python id": 10851, "finished": true, "Ev Idx": 13751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914246.267, "dur": 0.301,
    "args": {
      "Python parent id": 10848, "Python id": 10852, "finished": true, "Ev Idx": 13752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914246.816, "dur": 2.072,
    "args": {
      "Python parent id": 10848, "Python id": 10853, "finished": true, "Ev Idx": 13753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914249.575, "dur": 95.705,
    "args": {
      "Python parent id": 10848, "Python id": 10854, "finished": true, "Ev Idx": 13754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914250.639, "dur": 94.435,
    "args": {
      "Python parent id": 10854, "Python id": 10855, "finished": true, "Ev Idx": 13755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914251.020, "dur": 93.990,
    "args": {
      "Python parent id": 10855, "Python id": 10856, "finished": true, "Ev Idx": 13756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914257.795, "dur": 85.665,
    "args": {
      "Python parent id": 10856, "Python id": 10857, "finished": true, "Ev Idx": 13757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914258.142, "dur": 85.267,
    "args": {
      "Python parent id": 10857, "Python id": 10858, "finished": true, "Ev Idx": 13758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914258.792, "dur": 84.496,
    "args": {
      "Python parent id": 10858, "Python id": 10859, "finished": true, "Ev Idx": 13759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914259.412, "dur": 0.504,
    "args": {
      "Python parent id": 10859, "Python id": 10860, "finished": true, "Ev Idx": 13760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914259.721, "dur": 0.111,
    "args": {
      "Python parent id": 10860, "Python id": 10861, "finished": true, "Ev Idx": 13761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914260.328, "dur": 0.380,
    "args": {
      "Python parent id": 10859, "Python id": 10862, "finished": true, "Ev Idx": 13762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914260.611, "dur": 0.041,
    "args": {
      "Python parent id": 10862, "Python id": 10863, "finished": true, "Ev Idx": 13763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914261.097, "dur": 82.079,
    "args": {
      "Python parent id": 10859, "Python id": 10864, "finished": true, "Ev Idx": 13764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914261.916, "dur": 79.891,
    "args": {
      "Python parent id": 10864, "Python id": 10865, "finished": true, "Ev Idx": 13765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914270.305, "dur": 21.924,
    "args": {
      "Python parent id": 10865, "Python id": 10866, "finished": true, "Ev Idx": 13766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914270.467, "dur": 0.139,
    "args": {
      "Python parent id": 10866, "Python id": 10867, "finished": true, "Ev Idx": 13767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914270.992, "dur": 21.078,
    "args": {
      "Python parent id": 10866, "Python id": 10868, "finished": true, "Ev Idx": 13768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914271.144, "dur": 10.816,
    "args": {
      "Python parent id": 10868, "Python id": 10869, "finished": true, "Ev Idx": 13769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914300.111, "dur": 7.485,
    "args": {
      "Python parent id": 10865, "Python id": 10870, "finished": true, "Ev Idx": 13770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914315.980, "dur": 9.067,
    "args": {
      "Python parent id": 10865, "Python id": 10871, "finished": true, "Ev Idx": 13771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914346.689, "dur": 19.365,
    "args": {
      "Python parent id": 12, "Python id": 10872, "finished": true, "Ev Idx": 13772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914347.080, "dur": 14.564,
    "args": {
      "Python parent id": 10872, "Python id": 10873, "finished": true, "Ev Idx": 13773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914347.326, "dur": 0.589,
    "args": {
      "Python parent id": 10873, "Python id": 10874, "finished": true, "Ev Idx": 13774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914347.566, "dur": 0.256,
    "args": {
      "Python parent id": 10874, "Python id": 10875, "finished": true, "Ev Idx": 13775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914347.714, "dur": 0.074,
    "args": {
      "Python parent id": 10875, "Python id": 10876, "finished": true, "Ev Idx": 13776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914348.301, "dur": 10.449,
    "args": {
      "Python parent id": 10873, "Python id": 10877, "finished": true, "Ev Idx": 13777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914348.439, "dur": 0.093,
    "args": {
      "Python parent id": 10877, "Python id": 10878, "finished": true, "Ev Idx": 13778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914348.684, "dur": 0.065,
    "args": {
      "Python parent id": 10877, "Python id": 10879, "finished": true, "Ev Idx": 13779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914348.901, "dur": 0.063,
    "args": {
      "Python parent id": 10877, "Python id": 10880, "finished": true, "Ev Idx": 13780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914349.347, "dur": 0.041,
    "args": {
      "Python parent id": 10877, "Python id": 10881, "finished": true, "Ev Idx": 13781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914349.725, "dur": 0.098,
    "args": {
      "Python parent id": 10877, "Python id": 10882, "finished": true, "Ev Idx": 13782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914350.244, "dur": 8.453,
    "args": {
      "Python parent id": 10877, "Python id": 10883, "finished": true, "Ev Idx": 13783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914350.382, "dur": 0.063,
    "args": {
      "Python parent id": 10883, "Python id": 10884, "finished": true, "Ev Idx": 13784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914350.624, "dur": 0.058,
    "args": {
      "Python parent id": 10883, "Python id": 10885, "finished": true, "Ev Idx": 13785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914350.906, "dur": 0.066,
    "args": {
      "Python parent id": 10883, "Python id": 10886, "finished": true, "Ev Idx": 13786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914351.324, "dur": 0.036,
    "args": {
      "Python parent id": 10883, "Python id": 10887, "finished": true, "Ev Idx": 13787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914351.689, "dur": 6.885,
    "args": {
      "Python parent id": 10883, "Python id": 10888, "finished": true, "Ev Idx": 13788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914352.040, "dur": 6.423,
    "args": {
      "Python parent id": 10888, "Python id": 10889, "finished": true, "Ev Idx": 13789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914352.213, "dur": 4.188,
    "args": {
      "Python parent id": 10889, "Python id": 10890, "finished": true, "Ev Idx": 13790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914352.406, "dur": 3.919,
    "args": {
      "Python parent id": 10890, "Python id": 10891, "finished": true, "Ev Idx": 13791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914352.584, "dur": 0.294,
    "args": {
      "Python parent id": 10891, "Python id": 10892, "finished": true, "Ev Idx": 13792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914352.728, "dur": 0.098,
    "args": {
      "Python parent id": 10892, "Python id": 10893, "finished": true, "Ev Idx": 13793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914353.070, "dur": 2.830,
    "args": {
      "Python parent id": 10891, "Python id": 10894, "finished": true, "Ev Idx": 13794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914353.364, "dur": 2.413,
    "args": {
      "Python parent id": 10894, "Python id": 10895, "finished": true, "Ev Idx": 13795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914353.583, "dur": 2.115,
    "args": {
      "Python parent id": 10895, "Python id": 10896, "finished": true, "Ev Idx": 13796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914353.802, "dur": 1.572,
    "args": {
      "Python parent id": 10896, "Python id": 10897, "finished": true, "Ev Idx": 13797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914354.026, "dur": 0.675,
    "args": {
      "Python parent id": 10897, "Python id": 10898, "finished": true, "Ev Idx": 13798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914354.126, "dur": 0.050,
    "args": {
      "Python parent id": 10898, "Python id": 10899, "finished": true, "Ev Idx": 13799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914354.370, "dur": 0.254,
    "args": {
      "Python parent id": 10898, "Python id": 10900, "finished": true, "Ev Idx": 13800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914356.115, "dur": 0.177,
    "args": {
      "Python parent id": 10891, "Python id": 10901, "finished": true, "Ev Idx": 13801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914356.664, "dur": 0.142,
    "args": {
      "Python parent id": 10889, "Python id": 10902, "finished": true, "Ev Idx": 13802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914357.132, "dur": 1.284,
    "args": {
      "Python parent id": 10889, "Python id": 10903, "finished": true, "Ev Idx": 13803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914357.288, "dur": 1.070,
    "args": {
      "Python parent id": 10903, "Python id": 10904, "finished": true, "Ev Idx": 13804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914357.438, "dur": 0.432,
    "args": {
      "Python parent id": 10904, "Python id": 10905, "finished": true, "Ev Idx": 13805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914357.592, "dur": 0.223,
    "args": {
      "Python parent id": 10905, "Python id": 10906, "finished": true, "Ev Idx": 13806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914357.705, "dur": 0.057,
    "args": {
      "Python parent id": 10906, "Python id": 10907, "finished": true, "Ev Idx": 13807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914358.024, "dur": 0.290,
    "args": {
      "Python parent id": 10904, "Python id": 10908, "finished": true, "Ev Idx": 13808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914358.898, "dur": 0.298,
    "args": {
      "Python parent id": 10873, "Python id": 10909, "finished": true, "Ev Idx": 13809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914359.958, "dur": 1.479,
    "args": {
      "Python parent id": 10873, "Python id": 10910, "finished": true, "Ev Idx": 13810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914360.589, "dur": 0.756,
    "args": {
      "Python parent id": 10910, "Python id": 10911, "finished": true, "Ev Idx": 13811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914362.025, "dur": 3.935,
    "args": {
      "Python parent id": 10872, "Python id": 10912, "finished": true, "Ev Idx": 13812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914366.720, "dur": 9.044,
    "args": {
      "Python parent id": 12, "Python id": 10913, "finished": true, "Ev Idx": 13813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914376.253, "dur": 0.096,
    "args": {
      "Python parent id": 12, "Python id": 10914, "finished": true, "Ev Idx": 13814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914377.058, "dur": 0.230,
    "args": {
      "Python parent id": 12, "Python id": 10915, "finished": true, "Ev Idx": 13815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914377.805, "dur": 0.053,
    "args": {
      "Python parent id": 12, "Python id": 10916, "finished": true, "Ev Idx": 13816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914378.259, "dur": 10.529,
    "args": {
      "Python parent id": 12, "Python id": 10917, "finished": true, "Ev Idx": 13817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914390.170, "dur": 17.111,
    "args": {
      "Python parent id": 12, "Python id": 10918, "finished": true, "Ev Idx": 13818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914390.562, "dur": 12.372,
    "args": {
      "Python parent id": 10918, "Python id": 10919, "finished": true, "Ev Idx": 13819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914390.725, "dur": 0.592,
    "args": {
      "Python parent id": 10919, "Python id": 10920, "finished": true, "Ev Idx": 13820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914390.885, "dur": 0.343,
    "args": {
      "Python parent id": 10920, "Python id": 10921, "finished": true, "Ev Idx": 13821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914391.033, "dur": 0.127,
    "args": {
      "Python parent id": 10921, "Python id": 10922, "finished": true, "Ev Idx": 13822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914391.605, "dur": 9.142,
    "args": {
      "Python parent id": 10919, "Python id": 10923, "finished": true, "Ev Idx": 13823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914391.697, "dur": 0.106,
    "args": {
      "Python parent id": 10923, "Python id": 10924, "finished": true, "Ev Idx": 13824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914391.915, "dur": 0.063,
    "args": {
      "Python parent id": 10923, "Python id": 10925, "finished": true, "Ev Idx": 13825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914392.123, "dur": 0.062,
    "args": {
      "Python parent id": 10923, "Python id": 10926, "finished": true, "Ev Idx": 13826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914392.641, "dur": 0.090,
    "args": {
      "Python parent id": 10923, "Python id": 10927, "finished": true, "Ev Idx": 13827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914393.027, "dur": 0.070,
    "args": {
      "Python parent id": 10923, "Python id": 10928, "finished": true, "Ev Idx": 13828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914393.420, "dur": 7.274,
    "args": {
      "Python parent id": 10923, "Python id": 10929, "finished": true, "Ev Idx": 13829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914393.517, "dur": 0.061,
    "args": {
      "Python parent id": 10929, "Python id": 10930, "finished": true, "Ev Idx": 13830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914393.722, "dur": 0.055,
    "args": {
      "Python parent id": 10929, "Python id": 10931, "finished": true, "Ev Idx": 13831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914393.998, "dur": 0.074,
    "args": {
      "Python parent id": 10929, "Python id": 10932, "finished": true, "Ev Idx": 13832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914394.446, "dur": 0.063,
    "args": {
      "Python parent id": 10929, "Python id": 10933, "finished": true, "Ev Idx": 13833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914394.812, "dur": 5.754,
    "args": {
      "Python parent id": 10929, "Python id": 10934, "finished": true, "Ev Idx": 13834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.008, "dur": 5.478,
    "args": {
      "Python parent id": 10934, "Python id": 10935, "finished": true, "Ev Idx": 13835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.180, "dur": 3.341,
    "args": {
      "Python parent id": 10935, "Python id": 10936, "finished": true, "Ev Idx": 13836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.371, "dur": 3.102,
    "args": {
      "Python parent id": 10936, "Python id": 10937, "finished": true, "Ev Idx": 13837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.503, "dur": 0.283,
    "args": {
      "Python parent id": 10937, "Python id": 10938, "finished": true, "Ev Idx": 13838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.647, "dur": 0.068,
    "args": {
      "Python parent id": 10938, "Python id": 10939, "finished": true, "Ev Idx": 13839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914395.974, "dur": 2.096,
    "args": {
      "Python parent id": 10937, "Python id": 10940, "finished": true, "Ev Idx": 13840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914396.203, "dur": 1.778,
    "args": {
      "Python parent id": 10940, "Python id": 10941, "finished": true, "Ev Idx": 13841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914396.413, "dur": 1.520,
    "args": {
      "Python parent id": 10941, "Python id": 10942, "finished": true, "Ev Idx": 13842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914396.535, "dur": 1.193,
    "args": {
      "Python parent id": 10942, "Python id": 10943, "finished": true, "Ev Idx": 13843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914396.711, "dur": 0.505,
    "args": {
      "Python parent id": 10943, "Python id": 10944, "finished": true, "Ev Idx": 13844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914396.806, "dur": 0.052,
    "args": {
      "Python parent id": 10944, "Python id": 10945, "finished": true, "Ev Idx": 13845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914397.010, "dur": 0.144,
    "args": {
      "Python parent id": 10944, "Python id": 10946, "finished": true, "Ev Idx": 13846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914398.289, "dur": 0.145,
    "args": {
      "Python parent id": 10937, "Python id": 10947, "finished": true, "Ev Idx": 13847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914398.753, "dur": 0.142,
    "args": {
      "Python parent id": 10935, "Python id": 10948, "finished": true, "Ev Idx": 13848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914399.191, "dur": 1.252,
    "args": {
      "Python parent id": 10935, "Python id": 10949, "finished": true, "Ev Idx": 13849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914399.330, "dur": 1.068,
    "args": {
      "Python parent id": 10949, "Python id": 10950, "finished": true, "Ev Idx": 13850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914399.484, "dur": 0.399,
    "args": {
      "Python parent id": 10950, "Python id": 10951, "finished": true, "Ev Idx": 13851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914399.625, "dur": 0.201,
    "args": {
      "Python parent id": 10951, "Python id": 10952, "finished": true, "Ev Idx": 13852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914399.736, "dur": 0.056,
    "args": {
      "Python parent id": 10952, "Python id": 10953, "finished": true, "Ev Idx": 13853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914400.043, "dur": 0.309,
    "args": {
      "Python parent id": 10950, "Python id": 10954, "finished": true, "Ev Idx": 13854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914400.866, "dur": 0.181,
    "args": {
      "Python parent id": 10919, "Python id": 10955, "finished": true, "Ev Idx": 13855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914401.762, "dur": 1.039,
    "args": {
      "Python parent id": 10919, "Python id": 10956, "finished": true, "Ev Idx": 13856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914402.285, "dur": 0.465,
    "args": {
      "Python parent id": 10956, "Python id": 10957, "finished": true, "Ev Idx": 13857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914403.236, "dur": 3.969,
    "args": {
      "Python parent id": 10918, "Python id": 10958, "finished": true, "Ev Idx": 13858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914408.259, "dur": 93.892,
    "args": {
      "Python parent id": 12, "Python id": 10959, "finished": true, "Ev Idx": 13859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914408.627, "dur": 0.466,
    "args": {
      "Python parent id": 10959, "Python id": 10960, "finished": true, "Ev Idx": 13860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914408.935, "dur": 0.115,
    "args": {
      "Python parent id": 10960, "Python id": 10961, "finished": true, "Ev Idx": 13861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914409.410, "dur": 0.349,
    "args": {
      "Python parent id": 10959, "Python id": 10962, "finished": true, "Ev Idx": 13862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914410.031, "dur": 0.283,
    "args": {
      "Python parent id": 10959, "Python id": 10963, "finished": true, "Ev Idx": 13863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914410.572, "dur": 1.464,
    "args": {
      "Python parent id": 10959, "Python id": 10964, "finished": true, "Ev Idx": 13864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914412.741, "dur": 89.307,
    "args": {
      "Python parent id": 10959, "Python id": 10965, "finished": true, "Ev Idx": 13865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914413.646, "dur": 88.204,
    "args": {
      "Python parent id": 10965, "Python id": 10966, "finished": true, "Ev Idx": 13866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914413.982, "dur": 87.807,
    "args": {
      "Python parent id": 10966, "Python id": 10967, "finished": true, "Ev Idx": 13867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914419.661, "dur": 80.625,
    "args": {
      "Python parent id": 10967, "Python id": 10968, "finished": true, "Ev Idx": 13868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914420.033, "dur": 80.177,
    "args": {
      "Python parent id": 10968, "Python id": 10969, "finished": true, "Ev Idx": 13869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914420.652, "dur": 79.444,
    "args": {
      "Python parent id": 10969, "Python id": 10970, "finished": true, "Ev Idx": 13870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914421.163, "dur": 0.448,
    "args": {
      "Python parent id": 10970, "Python id": 10971, "finished": true, "Ev Idx": 13871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914421.467, "dur": 0.100,
    "args": {
      "Python parent id": 10971, "Python id": 10972, "finished": true, "Ev Idx": 13872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914421.962, "dur": 0.347,
    "args": {
      "Python parent id": 10970, "Python id": 10973, "finished": true, "Ev Idx": 13873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914422.217, "dur": 0.039,
    "args": {
      "Python parent id": 10973, "Python id": 10974, "finished": true, "Ev Idx": 13874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914422.708, "dur": 77.284,
    "args": {
      "Python parent id": 10970, "Python id": 10975, "finished": true, "Ev Idx": 13875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914423.464, "dur": 75.041,
    "args": {
      "Python parent id": 10975, "Python id": 10976, "finished": true, "Ev Idx": 13876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914430.798, "dur": 19.303,
    "args": {
      "Python parent id": 10976, "Python id": 10977, "finished": true, "Ev Idx": 13877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914430.957, "dur": 0.128,
    "args": {
      "Python parent id": 10977, "Python id": 10978, "finished": true, "Ev Idx": 13878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914431.404, "dur": 18.527,
    "args": {
      "Python parent id": 10977, "Python id": 10979, "finished": true, "Ev Idx": 13879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914431.578, "dur": 8.073,
    "args": {
      "Python parent id": 10979, "Python id": 10980, "finished": true, "Ev Idx": 13880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914457.613, "dur": 7.437,
    "args": {
      "Python parent id": 10976, "Python id": 10981, "finished": true, "Ev Idx": 13881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914474.401, "dur": 8.658,
    "args": {
      "Python parent id": 10976, "Python id": 10982, "finished": true, "Ev Idx": 13882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914503.459, "dur": 17.438,
    "args": {
      "Python parent id": 12, "Python id": 10983, "finished": true, "Ev Idx": 13883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914503.864, "dur": 13.448,
    "args": {
      "Python parent id": 10983, "Python id": 10984, "finished": true, "Ev Idx": 13884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914504.089, "dur": 0.481,
    "args": {
      "Python parent id": 10984, "Python id": 10985, "finished": true, "Ev Idx": 13885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914504.270, "dur": 0.234,
    "args": {
      "Python parent id": 10985, "Python id": 10986, "finished": true, "Ev Idx": 13886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914504.371, "dur": 0.099,
    "args": {
      "Python parent id": 10986, "Python id": 10987, "finished": true, "Ev Idx": 13887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914504.964, "dur": 9.820,
    "args": {
      "Python parent id": 10984, "Python id": 10988, "finished": true, "Ev Idx": 13888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914505.087, "dur": 0.068,
    "args": {
      "Python parent id": 10988, "Python id": 10989, "finished": true, "Ev Idx": 13889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914505.242, "dur": 0.069,
    "args": {
      "Python parent id": 10988, "Python id": 10990, "finished": true, "Ev Idx": 13890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914505.482, "dur": 0.067,
    "args": {
      "Python parent id": 10988, "Python id": 10991, "finished": true, "Ev Idx": 13891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914505.930, "dur": 0.043,
    "args": {
      "Python parent id": 10988, "Python id": 10992, "finished": true, "Ev Idx": 13892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914506.336, "dur": 0.101,
    "args": {
      "Python parent id": 10988, "Python id": 10993, "finished": true, "Ev Idx": 13893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914506.799, "dur": 7.904,
    "args": {
      "Python parent id": 10988, "Python id": 10994, "finished": true, "Ev Idx": 13894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914506.910, "dur": 0.071,
    "args": {
      "Python parent id": 10994, "Python id": 10995, "finished": true, "Ev Idx": 13895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914507.154, "dur": 0.085,
    "args": {
      "Python parent id": 10994, "Python id": 10996, "finished": true, "Ev Idx": 13896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914507.421, "dur": 0.063,
    "args": {
      "Python parent id": 10994, "Python id": 10997, "finished": true, "Ev Idx": 13897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914507.852, "dur": 0.037,
    "args": {
      "Python parent id": 10994, "Python id": 10998, "finished": true, "Ev Idx": 13898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914508.212, "dur": 6.394,
    "args": {
      "Python parent id": 10994, "Python id": 10999, "finished": true, "Ev Idx": 13899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914508.529, "dur": 6.019,
    "args": {
      "Python parent id": 10999, "Python id": 11000, "finished": true, "Ev Idx": 13900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914508.670, "dur": 3.893,
    "args": {
      "Python parent id": 11000, "Python id": 11001, "finished": true, "Ev Idx": 13901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914508.861, "dur": 3.621,
    "args": {
      "Python parent id": 11001, "Python id": 11002, "finished": true, "Ev Idx": 13902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914509.007, "dur": 0.250,
    "args": {
      "Python parent id": 11002, "Python id": 11003, "finished": true, "Ev Idx": 13903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914509.133, "dur": 0.061,
    "args": {
      "Python parent id": 11003, "Python id": 11004, "finished": true, "Ev Idx": 13904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914509.458, "dur": 2.616,
    "args": {
      "Python parent id": 11002, "Python id": 11005, "finished": true, "Ev Idx": 13905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914509.769, "dur": 2.189,
    "args": {
      "Python parent id": 11005, "Python id": 11006, "finished": true, "Ev Idx": 13906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914510.000, "dur": 1.889,
    "args": {
      "Python parent id": 11006, "Python id": 11007, "finished": true, "Ev Idx": 13907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914510.182, "dur": 1.426,
    "args": {
      "Python parent id": 11007, "Python id": 11008, "finished": true, "Ev Idx": 13908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914510.402, "dur": 0.573,
    "args": {
      "Python parent id": 11008, "Python id": 11009, "finished": true, "Ev Idx": 13909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914510.503, "dur": 0.042,
    "args": {
      "Python parent id": 11009, "Python id": 11010, "finished": true, "Ev Idx": 13910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914510.728, "dur": 0.202,
    "args": {
      "Python parent id": 11009, "Python id": 11011, "finished": true, "Ev Idx": 13911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914512.257, "dur": 0.180,
    "args": {
      "Python parent id": 11002, "Python id": 11012, "finished": true, "Ev Idx": 13912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914512.792, "dur": 0.166,
    "args": {
      "Python parent id": 11000, "Python id": 11013, "finished": true, "Ev Idx": 13913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914513.220, "dur": 1.284,
    "args": {
      "Python parent id": 11000, "Python id": 11014, "finished": true, "Ev Idx": 13914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914513.436, "dur": 1.007,
    "args": {
      "Python parent id": 11014, "Python id": 11015, "finished": true, "Ev Idx": 13915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914513.567, "dur": 0.417,
    "args": {
      "Python parent id": 11015, "Python id": 11016, "finished": true, "Ev Idx": 13916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914513.727, "dur": 0.196,
    "args": {
      "Python parent id": 11016, "Python id": 11017, "finished": true, "Ev Idx": 13917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914513.835, "dur": 0.057,
    "args": {
      "Python parent id": 11017, "Python id": 11018, "finished": true, "Ev Idx": 13918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914514.126, "dur": 0.275,
    "args": {
      "Python parent id": 11015, "Python id": 11019, "finished": true, "Ev Idx": 13919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914514.934, "dur": 0.264,
    "args": {
      "Python parent id": 10984, "Python id": 11020, "finished": true, "Ev Idx": 13920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914515.985, "dur": 1.158,
    "args": {
      "Python parent id": 10984, "Python id": 11021, "finished": true, "Ev Idx": 13921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914516.493, "dur": 0.570,
    "args": {
      "Python parent id": 11021, "Python id": 11022, "finished": true, "Ev Idx": 13922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914517.645, "dur": 3.197,
    "args": {
      "Python parent id": 10983, "Python id": 11023, "finished": true, "Ev Idx": 13923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914521.454, "dur": 7.917,
    "args": {
      "Python parent id": 12, "Python id": 11024, "finished": true, "Ev Idx": 13924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914529.852, "dur": 0.116,
    "args": {
      "Python parent id": 12, "Python id": 11025, "finished": true, "Ev Idx": 13925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914530.698, "dur": 0.221,
    "args": {
      "Python parent id": 12, "Python id": 11026, "finished": true, "Ev Idx": 13926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914531.534, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 11027, "finished": true, "Ev Idx": 13927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914532.036, "dur": 11.015,
    "args": {
      "Python parent id": 12, "Python id": 11028, "finished": true, "Ev Idx": 13928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914544.459, "dur": 15.465,
    "args": {
      "Python parent id": 12, "Python id": 11029, "finished": true, "Ev Idx": 13929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914544.803, "dur": 11.722,
    "args": {
      "Python parent id": 11029, "Python id": 11030, "finished": true, "Ev Idx": 13930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914545.011, "dur": 0.451,
    "args": {
      "Python parent id": 11030, "Python id": 11031, "finished": true, "Ev Idx": 13931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914545.162, "dur": 0.236,
    "args": {
      "Python parent id": 11031, "Python id": 11032, "finished": true, "Ev Idx": 13932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914545.262, "dur": 0.099,
    "args": {
      "Python parent id": 11032, "Python id": 11033, "finished": true, "Ev Idx": 13933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914545.752, "dur": 8.861,
    "args": {
      "Python parent id": 11030, "Python id": 11034, "finished": true, "Ev Idx": 13934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914545.844, "dur": 0.074,
    "args": {
      "Python parent id": 11034, "Python id": 11035, "finished": true, "Ev Idx": 13935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914546.028, "dur": 0.067,
    "args": {
      "Python parent id": 11034, "Python id": 11036, "finished": true, "Ev Idx": 13936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914546.266, "dur": 0.063,
    "args": {
      "Python parent id": 11034, "Python id": 11037, "finished": true, "Ev Idx": 13937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914546.721, "dur": 0.070,
    "args": {
      "Python parent id": 11034, "Python id": 11038, "finished": true, "Ev Idx": 13938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914547.085, "dur": 0.081,
    "args": {
      "Python parent id": 11034, "Python id": 11039, "finished": true, "Ev Idx": 13939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914547.582, "dur": 6.938,
    "args": {
      "Python parent id": 11034, "Python id": 11040, "finished": true, "Ev Idx": 13940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914547.680, "dur": 0.062,
    "args": {
      "Python parent id": 11040, "Python id": 11041, "finished": true, "Ev Idx": 13941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914547.896, "dur": 0.062,
    "args": {
      "Python parent id": 11040, "Python id": 11042, "finished": true, "Ev Idx": 13942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914548.161, "dur": 0.068,
    "args": {
      "Python parent id": 11040, "Python id": 11043, "finished": true, "Ev Idx": 13943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914548.690, "dur": 0.028,
    "args": {
      "Python parent id": 11040, "Python id": 11044, "finished": true, "Ev Idx": 13944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.041, "dur": 5.396,
    "args": {
      "Python parent id": 11040, "Python id": 11045, "finished": true, "Ev Idx": 13945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.314, "dur": 5.061,
    "args": {
      "Python parent id": 11045, "Python id": 11046, "finished": true, "Ev Idx": 13946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.492, "dur": 3.169,
    "args": {
      "Python parent id": 11046, "Python id": 11047, "finished": true, "Ev Idx": 13947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.684, "dur": 2.931,
    "args": {
      "Python parent id": 11047, "Python id": 11048, "finished": true, "Ev Idx": 13948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.841, "dur": 0.258,
    "args": {
      "Python parent id": 11048, "Python id": 11049, "finished": true, "Ev Idx": 13949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914549.981, "dur": 0.065,
    "args": {
      "Python parent id": 11049, "Python id": 11050, "finished": true, "Ev Idx": 13950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914550.239, "dur": 2.038,
    "args": {
      "Python parent id": 11048, "Python id": 11051, "finished": true, "Ev Idx": 13951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914550.447, "dur": 1.767,
    "args": {
      "Python parent id": 11051, "Python id": 11052, "finished": true, "Ev Idx": 13952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914550.651, "dur": 1.510,
    "args": {
      "Python parent id": 11052, "Python id": 11053, "finished": true, "Ev Idx": 13953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914550.849, "dur": 1.102,
    "args": {
      "Python parent id": 11053, "Python id": 11054, "finished": true, "Ev Idx": 13954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914551.021, "dur": 0.514,
    "args": {
      "Python parent id": 11054, "Python id": 11055, "finished": true, "Ev Idx": 13955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914551.107, "dur": 0.071,
    "args": {
      "Python parent id": 11055, "Python id": 11056, "finished": true, "Ev Idx": 13956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914551.356, "dur": 0.137,
    "args": {
      "Python parent id": 11055, "Python id": 11057, "finished": true, "Ev Idx": 13957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914552.484, "dur": 0.095,
    "args": {
      "Python parent id": 11048, "Python id": 11058, "finished": true, "Ev Idx": 13958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914552.825, "dur": 0.100,
    "args": {
      "Python parent id": 11046, "Python id": 11059, "finished": true, "Ev Idx": 13959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914553.229, "dur": 1.096,
    "args": {
      "Python parent id": 11046, "Python id": 11060, "finished": true, "Ev Idx": 13960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914553.366, "dur": 0.915,
    "args": {
      "Python parent id": 11060, "Python id": 11061, "finished": true, "Ev Idx": 13961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914553.521, "dur": 0.381,
    "args": {
      "Python parent id": 11061, "Python id": 11062, "finished": true, "Ev Idx": 13962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914553.662, "dur": 0.182,
    "args": {
      "Python parent id": 11062, "Python id": 11063, "finished": true, "Ev Idx": 13963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914553.752, "dur": 0.060,
    "args": {
      "Python parent id": 11063, "Python id": 11064, "finished": true, "Ev Idx": 13964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914554.035, "dur": 0.214,
    "args": {
      "Python parent id": 11061, "Python id": 11065, "finished": true, "Ev Idx": 13965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914554.758, "dur": 0.189,
    "args": {
      "Python parent id": 11030, "Python id": 11066, "finished": true, "Ev Idx": 13966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914555.507, "dur": 0.876,
    "args": {
      "Python parent id": 11030, "Python id": 11067, "finished": true, "Ev Idx": 13967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914555.903, "dur": 0.428,
    "args": {
      "Python parent id": 11067, "Python id": 11068, "finished": true, "Ev Idx": 13968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914556.849, "dur": 3.020,
    "args": {
      "Python parent id": 11029, "Python id": 11069, "finished": true, "Ev Idx": 13969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914560.898, "dur": 100.429,
    "args": {
      "Python parent id": 12, "Python id": 11070, "finished": true, "Ev Idx": 13970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914561.216, "dur": 0.439,
    "args": {
      "Python parent id": 11070, "Python id": 11071, "finished": true, "Ev Idx": 13971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914561.485, "dur": 0.086,
    "args": {
      "Python parent id": 11071, "Python id": 11072, "finished": true, "Ev Idx": 13972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914561.936, "dur": 0.315,
    "args": {
      "Python parent id": 11070, "Python id": 11073, "finished": true, "Ev Idx": 13973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914562.479, "dur": 0.300,
    "args": {
      "Python parent id": 11070, "Python id": 11074, "finished": true, "Ev Idx": 13974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914563.058, "dur": 1.483,
    "args": {
      "Python parent id": 11070, "Python id": 11075, "finished": true, "Ev Idx": 13975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914565.127, "dur": 96.092,
    "args": {
      "Python parent id": 11070, "Python id": 11076, "finished": true, "Ev Idx": 13976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914565.961, "dur": 95.063,
    "args": {
      "Python parent id": 11076, "Python id": 11077, "finished": true, "Ev Idx": 13977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914566.244, "dur": 94.696,
    "args": {
      "Python parent id": 11077, "Python id": 11078, "finished": true, "Ev Idx": 13978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914571.659, "dur": 87.902,
    "args": {
      "Python parent id": 11078, "Python id": 11079, "finished": true, "Ev Idx": 13979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914572.035, "dur": 87.438,
    "args": {
      "Python parent id": 11079, "Python id": 11080, "finished": true, "Ev Idx": 13980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914572.612, "dur": 86.734,
    "args": {
      "Python parent id": 11080, "Python id": 11081, "finished": true, "Ev Idx": 13981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914573.197, "dur": 0.395,
    "args": {
      "Python parent id": 11081, "Python id": 11082, "finished": true, "Ev Idx": 13982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914573.485, "dur": 0.045,
    "args": {
      "Python parent id": 11082, "Python id": 11083, "finished": true, "Ev Idx": 13983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914573.867, "dur": 0.399,
    "args": {
      "Python parent id": 11081, "Python id": 11084, "finished": true, "Ev Idx": 13984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914574.171, "dur": 0.039,
    "args": {
      "Python parent id": 11084, "Python id": 11085, "finished": true, "Ev Idx": 13985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914574.601, "dur": 84.630,
    "args": {
      "Python parent id": 11081, "Python id": 11086, "finished": true, "Ev Idx": 13986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914575.341, "dur": 82.605,
    "args": {
      "Python parent id": 11086, "Python id": 11087, "finished": true, "Ev Idx": 13987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914582.709, "dur": 18.090,
    "args": {
      "Python parent id": 11087, "Python id": 11088, "finished": true, "Ev Idx": 13988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914582.887, "dur": 0.141,
    "args": {
      "Python parent id": 11088, "Python id": 11089, "finished": true, "Ev Idx": 13989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914583.348, "dur": 17.264,
    "args": {
      "Python parent id": 11088, "Python id": 11090, "finished": true, "Ev Idx": 13990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914583.495, "dur": 7.357,
    "args": {
      "Python parent id": 11090, "Python id": 11091, "finished": true, "Ev Idx": 13991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914609.573, "dur": 7.427,
    "args": {
      "Python parent id": 11087, "Python id": 11092, "finished": true, "Ev Idx": 13992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914625.691, "dur": 8.258,
    "args": {
      "Python parent id": 11087, "Python id": 11093, "finished": true, "Ev Idx": 13993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914662.549, "dur": 17.111,
    "args": {
      "Python parent id": 12, "Python id": 11094, "finished": true, "Ev Idx": 13994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914663.047, "dur": 13.251,
    "args": {
      "Python parent id": 11094, "Python id": 11095, "finished": true, "Ev Idx": 13995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914663.298, "dur": 0.472,
    "args": {
      "Python parent id": 11095, "Python id": 11096, "finished": true, "Ev Idx": 13996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914663.448, "dur": 0.258,
    "args": {
      "Python parent id": 11096, "Python id": 11097, "finished": true, "Ev Idx": 13997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914663.585, "dur": 0.086,
    "args": {
      "Python parent id": 11097, "Python id": 11098, "finished": true, "Ev Idx": 13998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914664.090, "dur": 9.854,
    "args": {
      "Python parent id": 11095, "Python id": 11099, "finished": true, "Ev Idx": 13999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914664.177, "dur": 0.076,
    "args": {
      "Python parent id": 11099, "Python id": 11100, "finished": true, "Ev Idx": 14000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914664.378, "dur": 0.066,
    "args": {
      "Python parent id": 11099, "Python id": 11101, "finished": true, "Ev Idx": 14001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914664.596, "dur": 0.061,
    "args": {
      "Python parent id": 11099, "Python id": 11102, "finished": true, "Ev Idx": 14002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914665.190, "dur": 0.043,
    "args": {
      "Python parent id": 11099, "Python id": 11103, "finished": true, "Ev Idx": 14003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914665.626, "dur": 0.090,
    "args": {
      "Python parent id": 11099, "Python id": 11104, "finished": true, "Ev Idx": 14004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914666.080, "dur": 7.810,
    "args": {
      "Python parent id": 11099, "Python id": 11105, "finished": true, "Ev Idx": 14005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914666.186, "dur": 0.064,
    "args": {
      "Python parent id": 11105, "Python id": 11106, "finished": true, "Ev Idx": 14006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914666.414, "dur": 0.053,
    "args": {
      "Python parent id": 11105, "Python id": 11107, "finished": true, "Ev Idx": 14007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914666.652, "dur": 0.069,
    "args": {
      "Python parent id": 11105, "Python id": 11108, "finished": true, "Ev Idx": 14008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914667.147, "dur": 0.054,
    "args": {
      "Python parent id": 11105, "Python id": 11109, "finished": true, "Ev Idx": 14009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914667.559, "dur": 6.231,
    "args": {
      "Python parent id": 11105, "Python id": 11110, "finished": true, "Ev Idx": 14010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914667.850, "dur": 5.858,
    "args": {
      "Python parent id": 11110, "Python id": 11111, "finished": true, "Ev Idx": 14011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914667.972, "dur": 3.774,
    "args": {
      "Python parent id": 11111, "Python id": 11112, "finished": true, "Ev Idx": 14012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914668.198, "dur": 3.473,
    "args": {
      "Python parent id": 11112, "Python id": 11113, "finished": true, "Ev Idx": 14013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914668.347, "dur": 0.279,
    "args": {
      "Python parent id": 11113, "Python id": 11114, "finished": true, "Ev Idx": 14014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914668.502, "dur": 0.066,
    "args": {
      "Python parent id": 11114, "Python id": 11115, "finished": true, "Ev Idx": 14015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914668.768, "dur": 2.505,
    "args": {
      "Python parent id": 11113, "Python id": 11116, "finished": true, "Ev Idx": 14016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914669.014, "dur": 2.152,
    "args": {
      "Python parent id": 11116, "Python id": 11117, "finished": true, "Ev Idx": 14017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914669.208, "dur": 1.899,
    "args": {
      "Python parent id": 11117, "Python id": 11118, "finished": true, "Ev Idx": 14018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914669.480, "dur": 1.372,
    "args": {
      "Python parent id": 11118, "Python id": 11119, "finished": true, "Ev Idx": 14019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914669.682, "dur": 0.597,
    "args": {
      "Python parent id": 11119, "Python id": 11120, "finished": true, "Ev Idx": 14020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914669.777, "dur": 0.043,
    "args": {
      "Python parent id": 11120, "Python id": 11121, "finished": true, "Ev Idx": 14021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914670.021, "dur": 0.203,
    "args": {
      "Python parent id": 11120, "Python id": 11122, "finished": true, "Ev Idx": 14022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914671.466, "dur": 0.163,
    "args": {
      "Python parent id": 11113, "Python id": 11123, "finished": true, "Ev Idx": 14023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914671.972, "dur": 0.098,
    "args": {
      "Python parent id": 11111, "Python id": 11124, "finished": true, "Ev Idx": 14024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914672.374, "dur": 1.264,
    "args": {
      "Python parent id": 11111, "Python id": 11125, "finished": true, "Ev Idx": 14025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914672.529, "dur": 1.049,
    "args": {
      "Python parent id": 11125, "Python id": 11126, "finished": true, "Ev Idx": 14026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914672.669, "dur": 0.439,
    "args": {
      "Python parent id": 11126, "Python id": 11127, "finished": true, "Ev Idx": 14027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914672.829, "dur": 0.223,
    "args": {
      "Python parent id": 11127, "Python id": 11128, "finished": true, "Ev Idx": 14028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914672.937, "dur": 0.063,
    "args": {
      "Python parent id": 11128, "Python id": 11129, "finished": true, "Ev Idx": 14029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914673.253, "dur": 0.283,
    "args": {
      "Python parent id": 11126, "Python id": 11130, "finished": true, "Ev Idx": 14030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914674.077, "dur": 0.245,
    "args": {
      "Python parent id": 11095, "Python id": 11131, "finished": true, "Ev Idx": 14031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914674.989, "dur": 1.147,
    "args": {
      "Python parent id": 11095, "Python id": 11132, "finished": true, "Ev Idx": 14032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914675.555, "dur": 0.509,
    "args": {
      "Python parent id": 11132, "Python id": 11133, "finished": true, "Ev Idx": 14033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914676.602, "dur": 3.012,
    "args": {
      "Python parent id": 11094, "Python id": 11134, "finished": true, "Ev Idx": 14034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914680.226, "dur": 15.535,
    "args": {
      "Python parent id": 12, "Python id": 11135, "finished": true, "Ev Idx": 14035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914696.481, "dur": 0.135,
    "args": {
      "Python parent id": 12, "Python id": 11136, "finished": true, "Ev Idx": 14036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914697.606, "dur": 0.401,
    "args": {
      "Python parent id": 12, "Python id": 11137, "finished": true, "Ev Idx": 14037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914698.557, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 11138, "finished": true, "Ev Idx": 14038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914699.109, "dur": 15.844,
    "args": {
      "Python parent id": 12, "Python id": 11139, "finished": true, "Ev Idx": 14039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914716.842, "dur": 19.116,
    "args": {
      "Python parent id": 12, "Python id": 11140, "finished": true, "Ev Idx": 14040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914717.199, "dur": 14.196,
    "args": {
      "Python parent id": 11140, "Python id": 11141, "finished": true, "Ev Idx": 14041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914717.371, "dur": 0.659,
    "args": {
      "Python parent id": 11141, "Python id": 11142, "finished": true, "Ev Idx": 14042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914717.567, "dur": 0.330,
    "args": {
      "Python parent id": 11142, "Python id": 11143, "finished": true, "Ev Idx": 14043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914717.692, "dur": 0.132,
    "args": {
      "Python parent id": 11143, "Python id": 11144, "finished": true, "Ev Idx": 14044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914718.399, "dur": 10.210,
    "args": {
      "Python parent id": 11141, "Python id": 11145, "finished": true, "Ev Idx": 14045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914718.513, "dur": 0.090,
    "args": {
      "Python parent id": 11145, "Python id": 11146, "finished": true, "Ev Idx": 14046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914718.707, "dur": 0.064,
    "args": {
      "Python parent id": 11145, "Python id": 11147, "finished": true, "Ev Idx": 14047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914718.937, "dur": 0.067,
    "args": {
      "Python parent id": 11145, "Python id": 11148, "finished": true, "Ev Idx": 14048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914719.394, "dur": 0.037,
    "args": {
      "Python parent id": 11145, "Python id": 11149, "finished": true, "Ev Idx": 14049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914719.791, "dur": 0.078,
    "args": {
      "Python parent id": 11145, "Python id": 11150, "finished": true, "Ev Idx": 14050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914720.266, "dur": 8.287,
    "args": {
      "Python parent id": 11145, "Python id": 11151, "finished": true, "Ev Idx": 14051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914720.378, "dur": 0.067,
    "args": {
      "Python parent id": 11151, "Python id": 11152, "finished": true, "Ev Idx": 14052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914720.638, "dur": 0.052,
    "args": {
      "Python parent id": 11151, "Python id": 11153, "finished": true, "Ev Idx": 14053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914720.848, "dur": 0.074,
    "args": {
      "Python parent id": 11151, "Python id": 11154, "finished": true, "Ev Idx": 14054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914721.310, "dur": 0.053,
    "args": {
      "Python parent id": 11151, "Python id": 11155, "finished": true, "Ev Idx": 14055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914721.659, "dur": 6.769,
    "args": {
      "Python parent id": 11151, "Python id": 11156, "finished": true, "Ev Idx": 14056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914721.962, "dur": 6.384,
    "args": {
      "Python parent id": 11156, "Python id": 11157, "finished": true, "Ev Idx": 14057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914722.128, "dur": 4.161,
    "args": {
      "Python parent id": 11157, "Python id": 11158, "finished": true, "Ev Idx": 14058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914722.298, "dur": 3.928,
    "args": {
      "Python parent id": 11158, "Python id": 11159, "finished": true, "Ev Idx": 14059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914722.427, "dur": 0.331,
    "args": {
      "Python parent id": 11159, "Python id": 11160, "finished": true, "Ev Idx": 14060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914722.599, "dur": 0.091,
    "args": {
      "Python parent id": 11160, "Python id": 11161, "finished": true, "Ev Idx": 14061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914722.977, "dur": 2.881,
    "args": {
      "Python parent id": 11159, "Python id": 11162, "finished": true, "Ev Idx": 14062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914723.235, "dur": 2.480,
    "args": {
      "Python parent id": 11162, "Python id": 11163, "finished": true, "Ev Idx": 14063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914723.493, "dur": 2.181,
    "args": {
      "Python parent id": 11163, "Python id": 11164, "finished": true, "Ev Idx": 14064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914723.670, "dur": 1.640,
    "args": {
      "Python parent id": 11164, "Python id": 11165, "finished": true, "Ev Idx": 14065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914723.952, "dur": 0.595,
    "args": {
      "Python parent id": 11165, "Python id": 11166, "finished": true, "Ev Idx": 14066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914724.057, "dur": 0.064,
    "args": {
      "Python parent id": 11166, "Python id": 11167, "finished": true, "Ev Idx": 14067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914724.296, "dur": 0.191,
    "args": {
      "Python parent id": 11166, "Python id": 11168, "finished": true, "Ev Idx": 14068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914726.051, "dur": 0.113,
    "args": {
      "Python parent id": 11159, "Python id": 11169, "finished": true, "Ev Idx": 14069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914726.560, "dur": 0.104,
    "args": {
      "Python parent id": 11157, "Python id": 11170, "finished": true, "Ev Idx": 14070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914726.971, "dur": 1.333,
    "args": {
      "Python parent id": 11157, "Python id": 11171, "finished": true, "Ev Idx": 14071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914727.139, "dur": 1.107,
    "args": {
      "Python parent id": 11171, "Python id": 11172, "finished": true, "Ev Idx": 14072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914727.273, "dur": 0.413,
    "args": {
      "Python parent id": 11172, "Python id": 11173, "finished": true, "Ev Idx": 14073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914727.446, "dur": 0.177,
    "args": {
      "Python parent id": 11173, "Python id": 11174, "finished": true, "Ev Idx": 14074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914727.535, "dur": 0.059,
    "args": {
      "Python parent id": 11174, "Python id": 11175, "finished": true, "Ev Idx": 14075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914727.845, "dur": 0.342,
    "args": {
      "Python parent id": 11172, "Python id": 11176, "finished": true, "Ev Idx": 14076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914728.754, "dur": 0.271,
    "args": {
      "Python parent id": 11141, "Python id": 11177, "finished": true, "Ev Idx": 14077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914729.821, "dur": 1.409,
    "args": {
      "Python parent id": 11141, "Python id": 11178, "finished": true, "Ev Idx": 14078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914730.415, "dur": 0.705,
    "args": {
      "Python parent id": 11178, "Python id": 11179, "finished": true, "Ev Idx": 14079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914731.718, "dur": 4.162,
    "args": {
      "Python parent id": 11140, "Python id": 11180, "finished": true, "Ev Idx": 14080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914737.000, "dur": 112.402,
    "args": {
      "Python parent id": 12, "Python id": 11181, "finished": true, "Ev Idx": 14081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914737.421, "dur": 0.556,
    "args": {
      "Python parent id": 11181, "Python id": 11182, "finished": true, "Ev Idx": 14082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914737.802, "dur": 0.130,
    "args": {
      "Python parent id": 11182, "Python id": 11183, "finished": true, "Ev Idx": 14083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914738.315, "dur": 0.568,
    "args": {
      "Python parent id": 11181, "Python id": 11184, "finished": true, "Ev Idx": 14084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914739.115, "dur": 0.316,
    "args": {
      "Python parent id": 11181, "Python id": 11185, "finished": true, "Ev Idx": 14085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914739.685, "dur": 2.176,
    "args": {
      "Python parent id": 11181, "Python id": 11186, "finished": true, "Ev Idx": 14086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914742.583, "dur": 106.697,
    "args": {
      "Python parent id": 11181, "Python id": 11187, "finished": true, "Ev Idx": 14087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914743.639, "dur": 105.411,
    "args": {
      "Python parent id": 11187, "Python id": 11188, "finished": true, "Ev Idx": 14088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914744.020, "dur": 104.921,
    "args": {
      "Python parent id": 11188, "Python id": 11189, "finished": true, "Ev Idx": 14089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914751.043, "dur": 96.040,
    "args": {
      "Python parent id": 11189, "Python id": 11190, "finished": true, "Ev Idx": 14090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914751.447, "dur": 95.515,
    "args": {
      "Python parent id": 11190, "Python id": 11191, "finished": true, "Ev Idx": 14091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914752.068, "dur": 94.768,
    "args": {
      "Python parent id": 11191, "Python id": 11192, "finished": true, "Ev Idx": 14092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914752.653, "dur": 0.406,
    "args": {
      "Python parent id": 11192, "Python id": 11193, "finished": true, "Ev Idx": 14093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914752.939, "dur": 0.044,
    "args": {
      "Python parent id": 11193, "Python id": 11194, "finished": true, "Ev Idx": 14094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914753.422, "dur": 0.389,
    "args": {
      "Python parent id": 11192, "Python id": 11195, "finished": true, "Ev Idx": 14095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914753.712, "dur": 0.041,
    "args": {
      "Python parent id": 11195, "Python id": 11196, "finished": true, "Ev Idx": 14096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914754.194, "dur": 92.534,
    "args": {
      "Python parent id": 11192, "Python id": 11197, "finished": true, "Ev Idx": 14097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914754.931, "dur": 89.802,
    "args": {
      "Python parent id": 11197, "Python id": 11198, "finished": true, "Ev Idx": 14098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914764.609, "dur": 21.733,
    "args": {
      "Python parent id": 11198, "Python id": 11199, "finished": true, "Ev Idx": 14099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914764.794, "dur": 0.165,
    "args": {
      "Python parent id": 11199, "Python id": 11200, "finished": true, "Ev Idx": 14100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914765.293, "dur": 20.890,
    "args": {
      "Python parent id": 11199, "Python id": 11201, "finished": true, "Ev Idx": 14101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914765.395, "dur": 9.540,
    "args": {
      "Python parent id": 11201, "Python id": 11202, "finished": true, "Ev Idx": 14102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914794.593, "dur": 7.060,
    "args": {
      "Python parent id": 11198, "Python id": 11203, "finished": true, "Ev Idx": 14103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914811.350, "dur": 11.737,
    "args": {
      "Python parent id": 11198, "Python id": 11204, "finished": true, "Ev Idx": 14104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914851.070, "dur": 20.946,
    "args": {
      "Python parent id": 12, "Python id": 11205, "finished": true, "Ev Idx": 14105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914851.486, "dur": 16.011,
    "args": {
      "Python parent id": 11205, "Python id": 11206, "finished": true, "Ev Idx": 14106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914851.807, "dur": 0.746,
    "args": {
      "Python parent id": 11206, "Python id": 11207, "finished": true, "Ev Idx": 14107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914852.019, "dur": 0.427,
    "args": {
      "Python parent id": 11207, "Python id": 11208, "finished": true, "Ev Idx": 14108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914852.220, "dur": 0.157,
    "args": {
      "Python parent id": 11208, "Python id": 11209, "finished": true, "Ev Idx": 14109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914852.964, "dur": 11.306,
    "args": {
      "Python parent id": 11206, "Python id": 11210, "finished": true, "Ev Idx": 14110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914853.113, "dur": 0.091,
    "args": {
      "Python parent id": 11210, "Python id": 11211, "finished": true, "Ev Idx": 14111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914853.351, "dur": 0.059,
    "args": {
      "Python parent id": 11210, "Python id": 11212, "finished": true, "Ev Idx": 14112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914853.570, "dur": 0.061,
    "args": {
      "Python parent id": 11210, "Python id": 11213, "finished": true, "Ev Idx": 14113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914854.012, "dur": 0.056,
    "args": {
      "Python parent id": 11210, "Python id": 11214, "finished": true, "Ev Idx": 14114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914854.468, "dur": 0.088,
    "args": {
      "Python parent id": 11210, "Python id": 11215, "finished": true, "Ev Idx": 14115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914854.931, "dur": 9.285,
    "args": {
      "Python parent id": 11210, "Python id": 11216, "finished": true, "Ev Idx": 14116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914855.065, "dur": 0.066,
    "args": {
      "Python parent id": 11216, "Python id": 11217, "finished": true, "Ev Idx": 14117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914855.310, "dur": 0.066,
    "args": {
      "Python parent id": 11216, "Python id": 11218, "finished": true, "Ev Idx": 14118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914855.604, "dur": 0.090,
    "args": {
      "Python parent id": 11216, "Python id": 11219, "finished": true, "Ev Idx": 14119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914856.088, "dur": 0.072,
    "args": {
      "Python parent id": 11216, "Python id": 11220, "finished": true, "Ev Idx": 14120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914856.562, "dur": 7.526,
    "args": {
      "Python parent id": 11216, "Python id": 11221, "finished": true, "Ev Idx": 14121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914856.884, "dur": 7.117,
    "args": {
      "Python parent id": 11221, "Python id": 11222, "finished": true, "Ev Idx": 14122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914857.087, "dur": 4.630,
    "args": {
      "Python parent id": 11222, "Python id": 11223, "finished": true, "Ev Idx": 14123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914857.290, "dur": 4.351,
    "args": {
      "Python parent id": 11223, "Python id": 11224, "finished": true, "Ev Idx": 14124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914857.473, "dur": 0.347,
    "args": {
      "Python parent id": 11224, "Python id": 11225, "finished": true, "Ev Idx": 14125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914857.640, "dur": 0.094,
    "args": {
      "Python parent id": 11225, "Python id": 11226, "finished": true, "Ev Idx": 14126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914857.994, "dur": 3.171,
    "args": {
      "Python parent id": 11224, "Python id": 11227, "finished": true, "Ev Idx": 14127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914858.329, "dur": 2.711,
    "args": {
      "Python parent id": 11227, "Python id": 11228, "finished": true, "Ev Idx": 14128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914858.573, "dur": 2.409,
    "args": {
      "Python parent id": 11228, "Python id": 11229, "finished": true, "Ev Idx": 14129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914858.773, "dur": 1.843,
    "args": {
      "Python parent id": 11229, "Python id": 11230, "finished": true, "Ev Idx": 14130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914859.060, "dur": 0.694,
    "args": {
      "Python parent id": 11230, "Python id": 11231, "finished": true, "Ev Idx": 14131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914859.159, "dur": 0.066,
    "args": {
      "Python parent id": 11231, "Python id": 11232, "finished": true, "Ev Idx": 14132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914859.477, "dur": 0.232,
    "args": {
      "Python parent id": 11231, "Python id": 11233, "finished": true, "Ev Idx": 14133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914861.382, "dur": 0.218,
    "args": {
      "Python parent id": 11224, "Python id": 11234, "finished": true, "Ev Idx": 14134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914861.979, "dur": 0.136,
    "args": {
      "Python parent id": 11222, "Python id": 11235, "finished": true, "Ev Idx": 14135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914862.477, "dur": 1.482,
    "args": {
      "Python parent id": 11222, "Python id": 11236, "finished": true, "Ev Idx": 14136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914862.708, "dur": 1.194,
    "args": {
      "Python parent id": 11236, "Python id": 11237, "finished": true, "Ev Idx": 14137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914862.859, "dur": 0.465,
    "args": {
      "Python parent id": 11237, "Python id": 11238, "finished": true, "Ev Idx": 14138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914863.013, "dur": 0.234,
    "args": {
      "Python parent id": 11238, "Python id": 11239, "finished": true, "Ev Idx": 14139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914863.155, "dur": 0.058,
    "args": {
      "Python parent id": 11239, "Python id": 11240, "finished": true, "Ev Idx": 14140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914863.474, "dur": 0.387,
    "args": {
      "Python parent id": 11237, "Python id": 11241, "finished": true, "Ev Idx": 14141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914864.437, "dur": 0.390,
    "args": {
      "Python parent id": 11206, "Python id": 11242, "finished": true, "Ev Idx": 14142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914865.717, "dur": 1.584,
    "args": {
      "Python parent id": 11206, "Python id": 11243, "finished": true, "Ev Idx": 14143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914866.440, "dur": 0.783,
    "args": {
      "Python parent id": 11243, "Python id": 11244, "finished": true, "Ev Idx": 14144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914867.935, "dur": 4.005,
    "args": {
      "Python parent id": 11205, "Python id": 11245, "finished": true, "Ev Idx": 14145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914872.665, "dur": 10.078,
    "args": {
      "Python parent id": 12, "Python id": 11246, "finished": true, "Ev Idx": 14146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740914883.287, "dur": 0.113,
    "args": {
      "Python parent id": 12, "Python id": 11247, "finished": true, "Ev Idx": 14147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914884.195, "dur": 0.240,
    "args": {
      "Python parent id": 12, "Python id": 11248, "finished": true, "Ev Idx": 14148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914884.988, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 11249, "finished": true, "Ev Idx": 14149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740914885.467, "dur": 12.018,
    "args": {
      "Python parent id": 12, "Python id": 11250, "finished": true, "Ev Idx": 14150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740914898.946, "dur": 16.832,
    "args": {
      "Python parent id": 12, "Python id": 11251, "finished": true, "Ev Idx": 14151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740914899.283, "dur": 13.008,
    "args": {
      "Python parent id": 11251, "Python id": 11252, "finished": true, "Ev Idx": 14152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914899.436, "dur": 0.521,
    "args": {
      "Python parent id": 11252, "Python id": 11253, "finished": true, "Ev Idx": 14153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914899.618, "dur": 0.266,
    "args": {
      "Python parent id": 11253, "Python id": 11254, "finished": true, "Ev Idx": 14154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914899.754, "dur": 0.095,
    "args": {
      "Python parent id": 11254, "Python id": 11255, "finished": true, "Ev Idx": 14155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914900.343, "dur": 9.682,
    "args": {
      "Python parent id": 11252, "Python id": 11256, "finished": true, "Ev Idx": 14156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914900.429, "dur": 0.071,
    "args": {
      "Python parent id": 11256, "Python id": 11257, "finished": true, "Ev Idx": 14157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914900.605, "dur": 0.070,
    "args": {
      "Python parent id": 11256, "Python id": 11258, "finished": true, "Ev Idx": 14158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914900.818, "dur": 0.067,
    "args": {
      "Python parent id": 11256, "Python id": 11259, "finished": true, "Ev Idx": 14159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914901.243, "dur": 0.056,
    "args": {
      "Python parent id": 11256, "Python id": 11260, "finished": true, "Ev Idx": 14160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914901.599, "dur": 0.080,
    "args": {
      "Python parent id": 11256, "Python id": 11261, "finished": true, "Ev Idx": 14161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914902.140, "dur": 7.815,
    "args": {
      "Python parent id": 11256, "Python id": 11262, "finished": true, "Ev Idx": 14162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914902.238, "dur": 0.065,
    "args": {
      "Python parent id": 11262, "Python id": 11263, "finished": true, "Ev Idx": 14163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914902.475, "dur": 0.057,
    "args": {
      "Python parent id": 11262, "Python id": 11264, "finished": true, "Ev Idx": 14164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914902.715, "dur": 0.067,
    "args": {
      "Python parent id": 11262, "Python id": 11265, "finished": true, "Ev Idx": 14165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740914903.185, "dur": 0.048,
    "args": {
      "Python parent id": 11262, "Python id": 11266, "finished": true, "Ev Idx": 14166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740914903.559, "dur": 6.307,
    "args": {
      "Python parent id": 11262, "Python id": 11267, "finished": true, "Ev Idx": 14167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740914903.849, "dur": 5.932,
    "args": {
      "Python parent id": 11267, "Python id": 11268, "finished": true, "Ev Idx": 14168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740914904.011, "dur": 3.412,
    "args": {
      "Python parent id": 11268, "Python id": 11269, "finished": true, "Ev Idx": 14169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740914904.210, "dur": 3.135,
    "args": {
      "Python parent id": 11269, "Python id": 11270, "finished": true, "Ev Idx": 14170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740914904.346, "dur": 0.297,
    "args": {
      "Python parent id": 11270, "Python id": 11271, "finished": true, "Ev Idx": 14171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740914904.489, "dur": 0.087,
    "args": {
      "Python parent id": 11271, "Python id": 11272, "finished": true, "Ev Idx": 14172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740914904.843, "dur": 2.189,
    "args": {
      "Python parent id": 11270, "Python id": 11273, "finished": true, "Ev Idx": 14173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.038, "dur": 1.902,
    "args": {
      "Python parent id": 11273, "Python id": 11274, "finished": true, "Ev Idx": 14174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.265, "dur": 1.632,
    "args": {
      "Python parent id": 11274, "Python id": 11275, "finished": true, "Ev Idx": 14175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.416, "dur": 1.222,
    "args": {
      "Python parent id": 11275, "Python id": 11276, "finished": true, "Ev Idx": 14176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.608, "dur": 0.530,
    "args": {
      "Python parent id": 11276, "Python id": 11277, "finished": true, "Ev Idx": 14177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.708, "dur": 0.046,
    "args": {
      "Python parent id": 11277, "Python id": 11278, "finished": true, "Ev Idx": 14178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740914905.943, "dur": 0.150,
    "args": {
      "Python parent id": 11277, "Python id": 11279, "finished": true, "Ev Idx": 14179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740914907.209, "dur": 0.103,
    "args": {
      "Python parent id": 11270, "Python id": 11280, "finished": true, "Ev Idx": 14180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740914907.636, "dur": 0.122,
    "args": {
      "Python parent id": 11268, "Python id": 11281, "finished": true, "Ev Idx": 14181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740914908.028, "dur": 1.705,
    "args": {
      "Python parent id": 11268, "Python id": 11282, "finished": true, "Ev Idx": 14182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740914908.631, "dur": 1.056,
    "args": {
      "Python parent id": 11282, "Python id": 11283, "finished": true, "Ev Idx": 14183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740914908.811, "dur": 0.427,
    "args": {
      "Python parent id": 11283, "Python id": 11284, "finished": true, "Ev Idx": 14184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740914908.972, "dur": 0.188,
    "args": {
      "Python parent id": 11284, "Python id": 11285, "finished": true, "Ev Idx": 14185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740914909.061, "dur": 0.062,
    "args": {
      "Python parent id": 11285, "Python id": 11286, "finished": true, "Ev Idx": 14186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740914909.383, "dur": 0.250,
    "args": {
      "Python parent id": 11283, "Python id": 11287, "finished": true, "Ev Idx": 14187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740914910.183, "dur": 0.210,
    "args": {
      "Python parent id": 11252, "Python id": 11288, "finished": true, "Ev Idx": 14188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740914911.123, "dur": 1.030,
    "args": {
      "Python parent id": 11252, "Python id": 11289, "finished": true, "Ev Idx": 14189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740914911.631, "dur": 0.474,
    "args": {
      "Python parent id": 11289, "Python id": 11290, "finished": true, "Ev Idx": 14190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740914912.604, "dur": 3.116,
    "args": {
      "Python parent id": 11251, "Python id": 11291, "finished": true, "Ev Idx": 14191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740914916.793, "dur": 101.724,
    "args": {
      "Python parent id": 12, "Python id": 11292, "finished": true, "Ev Idx": 14192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914917.185, "dur": 0.519,
    "args": {
      "Python parent id": 11292, "Python id": 11293, "finished": true, "Ev Idx": 14193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914917.525, "dur": 0.116,
    "args": {
      "Python parent id": 11293, "Python id": 11294, "finished": true, "Ev Idx": 14194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914918.015, "dur": 0.529,
    "args": {
      "Python parent id": 11292, "Python id": 11295, "finished": true, "Ev Idx": 14195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740914918.804, "dur": 0.280,
    "args": {
      "Python parent id": 11292, "Python id": 11296, "finished": true, "Ev Idx": 14196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740914919.355, "dur": 1.814,
    "args": {
      "Python parent id": 11292, "Python id": 11297, "finished": true, "Ev Idx": 14197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740914921.803, "dur": 96.607,
    "args": {
      "Python parent id": 11292, "Python id": 11298, "finished": true, "Ev Idx": 14198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740914922.802, "dur": 95.382,
    "args": {
      "Python parent id": 11298, "Python id": 11299, "finished": true, "Ev Idx": 14199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914923.176, "dur": 94.946,
    "args": {
      "Python parent id": 11299, "Python id": 11300, "finished": true, "Ev Idx": 14200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740914929.849, "dur": 86.822,
    "args": {
      "Python parent id": 11300, "Python id": 11301, "finished": true, "Ev Idx": 14201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740914930.179, "dur": 86.389,
    "args": {
      "Python parent id": 11301, "Python id": 11302, "finished": true, "Ev Idx": 14202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740914930.788, "dur": 85.625,
    "args": {
      "Python parent id": 11302, "Python id": 11303, "finished": true, "Ev Idx": 14203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914931.439, "dur": 0.541,
    "args": {
      "Python parent id": 11303, "Python id": 11304, "finished": true, "Ev Idx": 14204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914931.802, "dur": 0.112,
    "args": {
      "Python parent id": 11304, "Python id": 11305, "finished": true, "Ev Idx": 14205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740914932.287, "dur": 0.336,
    "args": {
      "Python parent id": 11303, "Python id": 11306, "finished": true, "Ev Idx": 14206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740914932.520, "dur": 0.049,
    "args": {
      "Python parent id": 11306, "Python id": 11307, "finished": true, "Ev Idx": 14207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914932.989, "dur": 83.308,
    "args": {
      "Python parent id": 11303, "Python id": 11308, "finished": true, "Ev Idx": 14208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740914933.802, "dur": 80.996,
    "args": {
      "Python parent id": 11308, "Python id": 11309, "finished": true, "Ev Idx": 14209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740914942.019, "dur": 21.881,
    "args": {
      "Python parent id": 11309, "Python id": 11310, "finished": true, "Ev Idx": 14210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740914942.190, "dur": 0.133,
    "args": {
      "Python parent id": 11310, "Python id": 11311, "finished": true, "Ev Idx": 14211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740914942.708, "dur": 21.047,
    "args": {
      "Python parent id": 11310, "Python id": 11312, "finished": true, "Ev Idx": 14212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740914942.863, "dur": 9.469,
    "args": {
      "Python parent id": 11312, "Python id": 11313, "finished": true, "Ev Idx": 14213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914971.878, "dur": 8.214,
    "args": {
      "Python parent id": 11309, "Python id": 11314, "finished": true, "Ev Idx": 14214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740914989.782, "dur": 9.106,
    "args": {
      "Python parent id": 11309, "Python id": 11315, "finished": true, "Ev Idx": 14215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915019.909, "dur": 19.385,
    "args": {
      "Python parent id": 12, "Python id": 11316, "finished": true, "Ev Idx": 14216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915020.316, "dur": 14.478,
    "args": {
      "Python parent id": 11316, "Python id": 11317, "finished": true, "Ev Idx": 14217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915020.503, "dur": 0.676,
    "args": {
      "Python parent id": 11317, "Python id": 11318, "finished": true, "Ev Idx": 14218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915020.773, "dur": 0.336,
    "args": {
      "Python parent id": 11318, "Python id": 11319, "finished": true, "Ev Idx": 14219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915020.928, "dur": 0.146,
    "args": {
      "Python parent id": 11319, "Python id": 11320, "finished": true, "Ev Idx": 14220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915021.551, "dur": 10.458,
    "args": {
      "Python parent id": 11317, "Python id": 11321, "finished": true, "Ev Idx": 14221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915021.674, "dur": 0.097,
    "args": {
      "Python parent id": 11321, "Python id": 11322, "finished": true, "Ev Idx": 14222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915021.945, "dur": 0.081,
    "args": {
      "Python parent id": 11321, "Python id": 11323, "finished": true, "Ev Idx": 14223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915022.206, "dur": 0.085,
    "args": {
      "Python parent id": 11321, "Python id": 11324, "finished": true, "Ev Idx": 14224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915022.705, "dur": 0.037,
    "args": {
      "Python parent id": 11321, "Python id": 11325, "finished": true, "Ev Idx": 14225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915023.046, "dur": 0.100,
    "args": {
      "Python parent id": 11321, "Python id": 11326, "finished": true, "Ev Idx": 14226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915023.586, "dur": 8.356,
    "args": {
      "Python parent id": 11321, "Python id": 11327, "finished": true, "Ev Idx": 14227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915023.675, "dur": 0.061,
    "args": {
      "Python parent id": 11327, "Python id": 11328, "finished": true, "Ev Idx": 14228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915023.945, "dur": 0.054,
    "args": {
      "Python parent id": 11327, "Python id": 11329, "finished": true, "Ev Idx": 14229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915024.204, "dur": 0.063,
    "args": {
      "Python parent id": 11327, "Python id": 11330, "finished": true, "Ev Idx": 14230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915024.611, "dur": 0.055,
    "args": {
      "Python parent id": 11327, "Python id": 11331, "finished": true, "Ev Idx": 14231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915024.991, "dur": 6.845,
    "args": {
      "Python parent id": 11327, "Python id": 11332, "finished": true, "Ev Idx": 14232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915025.305, "dur": 6.446,
    "args": {
      "Python parent id": 11332, "Python id": 11333, "finished": true, "Ev Idx": 14233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915025.438, "dur": 4.152,
    "args": {
      "Python parent id": 11333, "Python id": 11334, "finished": true, "Ev Idx": 14234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915025.636, "dur": 3.862,
    "args": {
      "Python parent id": 11334, "Python id": 11335, "finished": true, "Ev Idx": 14235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915025.768, "dur": 0.293,
    "args": {
      "Python parent id": 11335, "Python id": 11336, "finished": true, "Ev Idx": 14236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915025.917, "dur": 0.066,
    "args": {
      "Python parent id": 11336, "Python id": 11337, "finished": true, "Ev Idx": 14237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915026.268, "dur": 2.827,
    "args": {
      "Python parent id": 11335, "Python id": 11338, "finished": true, "Ev Idx": 14238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915026.602, "dur": 2.367,
    "args": {
      "Python parent id": 11338, "Python id": 11339, "finished": true, "Ev Idx": 14239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915026.884, "dur": 2.018,
    "args": {
      "Python parent id": 11339, "Python id": 11340, "finished": true, "Ev Idx": 14240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915027.065, "dur": 1.525,
    "args": {
      "Python parent id": 11340, "Python id": 11341, "finished": true, "Ev Idx": 14241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915027.288, "dur": 0.669,
    "args": {
      "Python parent id": 11341, "Python id": 11342, "finished": true, "Ev Idx": 14242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915027.395, "dur": 0.044,
    "args": {
      "Python parent id": 11342, "Python id": 11343, "finished": true, "Ev Idx": 14243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915027.674, "dur": 0.221,
    "args": {
      "Python parent id": 11342, "Python id": 11344, "finished": true, "Ev Idx": 14244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915029.287, "dur": 0.175,
    "args": {
      "Python parent id": 11335, "Python id": 11345, "finished": true, "Ev Idx": 14245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915029.831, "dur": 0.175,
    "args": {
      "Python parent id": 11333, "Python id": 11346, "finished": true, "Ev Idx": 14246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915030.318, "dur": 1.371,
    "args": {
      "Python parent id": 11333, "Python id": 11347, "finished": true, "Ev Idx": 14247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915030.505, "dur": 1.127,
    "args": {
      "Python parent id": 11347, "Python id": 11348, "finished": true, "Ev Idx": 14248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915030.655, "dur": 0.488,
    "args": {
      "Python parent id": 11348, "Python id": 11349, "finished": true, "Ev Idx": 14249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915030.814, "dur": 0.252,
    "args": {
      "Python parent id": 11349, "Python id": 11350, "finished": true, "Ev Idx": 14250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915030.961, "dur": 0.074,
    "args": {
      "Python parent id": 11350, "Python id": 11351, "finished": true, "Ev Idx": 14251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915031.267, "dur": 0.324,
    "args": {
      "Python parent id": 11348, "Python id": 11352, "finished": true, "Ev Idx": 14252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915032.178, "dur": 0.348,
    "args": {
      "Python parent id": 11317, "Python id": 11353, "finished": true, "Ev Idx": 14253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915033.289, "dur": 1.340,
    "args": {
      "Python parent id": 11317, "Python id": 11354, "finished": true, "Ev Idx": 14254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915033.919, "dur": 0.636,
    "args": {
      "Python parent id": 11354, "Python id": 11355, "finished": true, "Ev Idx": 14255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915035.199, "dur": 4.000,
    "args": {
      "Python parent id": 11316, "Python id": 11356, "finished": true, "Ev Idx": 14256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915039.943, "dur": 7.764,
    "args": {
      "Python parent id": 12, "Python id": 11357, "finished": true, "Ev Idx": 14257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915048.180, "dur": 0.129,
    "args": {
      "Python parent id": 12, "Python id": 11358, "finished": true, "Ev Idx": 14258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915049.097, "dur": 0.202,
    "args": {
      "Python parent id": 12, "Python id": 11359, "finished": true, "Ev Idx": 14259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915049.809, "dur": 0.075,
    "args": {
      "Python parent id": 12, "Python id": 11360, "finished": true, "Ev Idx": 14260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915050.300, "dur": 10.864,
    "args": {
      "Python parent id": 12, "Python id": 11361, "finished": true, "Ev Idx": 14261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915062.467, "dur": 15.899,
    "args": {
      "Python parent id": 12, "Python id": 11362, "finished": true, "Ev Idx": 14262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915062.788, "dur": 11.976,
    "args": {
      "Python parent id": 11362, "Python id": 11363, "finished": true, "Ev Idx": 14263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915062.941, "dur": 0.559,
    "args": {
      "Python parent id": 11363, "Python id": 11364, "finished": true, "Ev Idx": 14264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915063.098, "dur": 0.287,
    "args": {
      "Python parent id": 11364, "Python id": 11365, "finished": true, "Ev Idx": 14265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915063.199, "dur": 0.149,
    "args": {
      "Python parent id": 11365, "Python id": 11366, "finished": true, "Ev Idx": 14266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915063.767, "dur": 9.011,
    "args": {
      "Python parent id": 11363, "Python id": 11367, "finished": true, "Ev Idx": 14267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915063.848, "dur": 0.108,
    "args": {
      "Python parent id": 11367, "Python id": 11368, "finished": true, "Ev Idx": 14268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915064.065, "dur": 0.066,
    "args": {
      "Python parent id": 11367, "Python id": 11369, "finished": true, "Ev Idx": 14269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915064.287, "dur": 0.062,
    "args": {
      "Python parent id": 11367, "Python id": 11370, "finished": true, "Ev Idx": 14270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915064.748, "dur": 0.074,
    "args": {
      "Python parent id": 11367, "Python id": 11371, "finished": true, "Ev Idx": 14271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915065.070, "dur": 0.077,
    "args": {
      "Python parent id": 11367, "Python id": 11372, "finished": true, "Ev Idx": 14272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915065.593, "dur": 7.117,
    "args": {
      "Python parent id": 11367, "Python id": 11373, "finished": true, "Ev Idx": 14273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915065.702, "dur": 0.071,
    "args": {
      "Python parent id": 11373, "Python id": 11374, "finished": true, "Ev Idx": 14274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915065.918, "dur": 0.061,
    "args": {
      "Python parent id": 11373, "Python id": 11375, "finished": true, "Ev Idx": 14275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915066.163, "dur": 0.073,
    "args": {
      "Python parent id": 11373, "Python id": 11376, "finished": true, "Ev Idx": 14276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915066.619, "dur": 0.040,
    "args": {
      "Python parent id": 11373, "Python id": 11377, "finished": true, "Ev Idx": 14277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915066.995, "dur": 5.620,
    "args": {
      "Python parent id": 11373, "Python id": 11378, "finished": true, "Ev Idx": 14278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915067.192, "dur": 5.364,
    "args": {
      "Python parent id": 11378, "Python id": 11379, "finished": true, "Ev Idx": 14279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915067.327, "dur": 3.409,
    "args": {
      "Python parent id": 11379, "Python id": 11380, "finished": true, "Ev Idx": 14280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915067.522, "dur": 3.147,
    "args": {
      "Python parent id": 11380, "Python id": 11381, "finished": true, "Ev Idx": 14281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915067.645, "dur": 0.306,
    "args": {
      "Python parent id": 11381, "Python id": 11382, "finished": true, "Ev Idx": 14282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915067.787, "dur": 0.066,
    "args": {
      "Python parent id": 11382, "Python id": 11383, "finished": true, "Ev Idx": 14283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915068.123, "dur": 2.248,
    "args": {
      "Python parent id": 11381, "Python id": 11384, "finished": true, "Ev Idx": 14284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915068.337, "dur": 1.945,
    "args": {
      "Python parent id": 11384, "Python id": 11385, "finished": true, "Ev Idx": 14285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915068.605, "dur": 1.633,
    "args": {
      "Python parent id": 11385, "Python id": 11386, "finished": true, "Ev Idx": 14286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915068.791, "dur": 1.215,
    "args": {
      "Python parent id": 11386, "Python id": 11387, "finished": true, "Ev Idx": 14287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915069.009, "dur": 0.545,
    "args": {
      "Python parent id": 11387, "Python id": 11388, "finished": true, "Ev Idx": 14288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915069.107, "dur": 0.075,
    "args": {
      "Python parent id": 11388, "Python id": 11389, "finished": true, "Ev Idx": 14289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915069.360, "dur": 0.151,
    "args": {
      "Python parent id": 11388, "Python id": 11390, "finished": true, "Ev Idx": 14290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915070.552, "dur": 0.068,
    "args": {
      "Python parent id": 11381, "Python id": 11391, "finished": true, "Ev Idx": 14291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915070.918, "dur": 0.096,
    "args": {
      "Python parent id": 11379, "Python id": 11392, "finished": true, "Ev Idx": 14292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915071.313, "dur": 1.180,
    "args": {
      "Python parent id": 11379, "Python id": 11393, "finished": true, "Ev Idx": 14293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915071.463, "dur": 0.986,
    "args": {
      "Python parent id": 11393, "Python id": 11394, "finished": true, "Ev Idx": 14294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915071.597, "dur": 0.417,
    "args": {
      "Python parent id": 11394, "Python id": 11395, "finished": true, "Ev Idx": 14295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915071.748, "dur": 0.194,
    "args": {
      "Python parent id": 11395, "Python id": 11396, "finished": true, "Ev Idx": 14296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915071.852, "dur": 0.057,
    "args": {
      "Python parent id": 11396, "Python id": 11397, "finished": true, "Ev Idx": 14297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915072.160, "dur": 0.260,
    "args": {
      "Python parent id": 11394, "Python id": 11398, "finished": true, "Ev Idx": 14298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915072.927, "dur": 0.146,
    "args": {
      "Python parent id": 11363, "Python id": 11399, "finished": true, "Ev Idx": 14299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915073.672, "dur": 0.951,
    "args": {
      "Python parent id": 11363, "Python id": 11400, "finished": true, "Ev Idx": 14300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915074.128, "dur": 0.445,
    "args": {
      "Python parent id": 11400, "Python id": 11401, "finished": true, "Ev Idx": 14301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915075.058, "dur": 3.248,
    "args": {
      "Python parent id": 11362, "Python id": 11402, "finished": true, "Ev Idx": 14302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915079.347, "dur": 94.768,
    "args": {
      "Python parent id": 12, "Python id": 11403, "finished": true, "Ev Idx": 14303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915079.702, "dur": 0.467,
    "args": {
      "Python parent id": 11403, "Python id": 11404, "finished": true, "Ev Idx": 14304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915080.004, "dur": 0.081,
    "args": {
      "Python parent id": 11404, "Python id": 11405, "finished": true, "Ev Idx": 14305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915080.427, "dur": 0.352,
    "args": {
      "Python parent id": 11403, "Python id": 11406, "finished": true, "Ev Idx": 14306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915081.031, "dur": 0.262,
    "args": {
      "Python parent id": 11403, "Python id": 11407, "finished": true, "Ev Idx": 14307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915081.526, "dur": 1.756,
    "args": {
      "Python parent id": 11403, "Python id": 11408, "finished": true, "Ev Idx": 14308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915083.998, "dur": 90.047,
    "args": {
      "Python parent id": 11403, "Python id": 11409, "finished": true, "Ev Idx": 14309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915084.909, "dur": 88.924,
    "args": {
      "Python parent id": 11409, "Python id": 11410, "finished": true, "Ev Idx": 14310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915085.231, "dur": 88.516,
    "args": {
      "Python parent id": 11410, "Python id": 11411, "finished": true, "Ev Idx": 14311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915090.890, "dur": 81.433,
    "args": {
      "Python parent id": 11411, "Python id": 11412, "finished": true, "Ev Idx": 14312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915091.273, "dur": 80.939,
    "args": {
      "Python parent id": 11412, "Python id": 11413, "finished": true, "Ev Idx": 14313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915091.876, "dur": 80.152,
    "args": {
      "Python parent id": 11413, "Python id": 11414, "finished": true, "Ev Idx": 14314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915092.456, "dur": 0.499,
    "args": {
      "Python parent id": 11414, "Python id": 11415, "finished": true, "Ev Idx": 14315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915092.802, "dur": 0.047,
    "args": {
      "Python parent id": 11415, "Python id": 11416, "finished": true, "Ev Idx": 14316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915093.257, "dur": 0.341,
    "args": {
      "Python parent id": 11414, "Python id": 11417, "finished": true, "Ev Idx": 14317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915093.501, "dur": 0.038,
    "args": {
      "Python parent id": 11417, "Python id": 11418, "finished": true, "Ev Idx": 14318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915093.967, "dur": 77.931,
    "args": {
      "Python parent id": 11414, "Python id": 11419, "finished": true, "Ev Idx": 14319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915094.687, "dur": 75.858,
    "args": {
      "Python parent id": 11419, "Python id": 11420, "finished": true, "Ev Idx": 14320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915102.085, "dur": 17.921,
    "args": {
      "Python parent id": 11420, "Python id": 11421, "finished": true, "Ev Idx": 14321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915102.251, "dur": 0.127,
    "args": {
      "Python parent id": 11421, "Python id": 11422, "finished": true, "Ev Idx": 14322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915102.733, "dur": 17.124,
    "args": {
      "Python parent id": 11421, "Python id": 11423, "finished": true, "Ev Idx": 14323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915102.835, "dur": 7.260,
    "args": {
      "Python parent id": 11423, "Python id": 11424, "finished": true, "Ev Idx": 14324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915128.262, "dur": 7.278,
    "args": {
      "Python parent id": 11420, "Python id": 11425, "finished": true, "Ev Idx": 14325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915145.192, "dur": 8.617,
    "args": {
      "Python parent id": 11420, "Python id": 11426, "finished": true, "Ev Idx": 14326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915175.510, "dur": 17.697,
    "args": {
      "Python parent id": 12, "Python id": 11427, "finished": true, "Ev Idx": 14327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915175.833, "dur": 13.611,
    "args": {
      "Python parent id": 11427, "Python id": 11428, "finished": true, "Ev Idx": 14328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915176.003, "dur": 0.612,
    "args": {
      "Python parent id": 11428, "Python id": 11429, "finished": true, "Ev Idx": 14329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915176.175, "dur": 0.341,
    "args": {
      "Python parent id": 11429, "Python id": 11430, "finished": true, "Ev Idx": 14330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915176.332, "dur": 0.129,
    "args": {
      "Python parent id": 11430, "Python id": 11431, "finished": true, "Ev Idx": 14331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915177.006, "dur": 9.833,
    "args": {
      "Python parent id": 11428, "Python id": 11432, "finished": true, "Ev Idx": 14332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915177.121, "dur": 0.064,
    "args": {
      "Python parent id": 11432, "Python id": 11433, "finished": true, "Ev Idx": 14333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915177.325, "dur": 0.062,
    "args": {
      "Python parent id": 11432, "Python id": 11434, "finished": true, "Ev Idx": 14334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915177.540, "dur": 0.058,
    "args": {
      "Python parent id": 11432, "Python id": 11435, "finished": true, "Ev Idx": 14335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915178.037, "dur": 0.076,
    "args": {
      "Python parent id": 11432, "Python id": 11436, "finished": true, "Ev Idx": 14336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915178.417, "dur": 0.076,
    "args": {
      "Python parent id": 11432, "Python id": 11437, "finished": true, "Ev Idx": 14337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915178.822, "dur": 7.945,
    "args": {
      "Python parent id": 11432, "Python id": 11438, "finished": true, "Ev Idx": 14338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915178.927, "dur": 0.065,
    "args": {
      "Python parent id": 11438, "Python id": 11439, "finished": true, "Ev Idx": 14339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915179.163, "dur": 0.059,
    "args": {
      "Python parent id": 11438, "Python id": 11440, "finished": true, "Ev Idx": 14340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915179.404, "dur": 0.073,
    "args": {
      "Python parent id": 11438, "Python id": 11441, "finished": true, "Ev Idx": 14341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915179.805, "dur": 0.034,
    "args": {
      "Python parent id": 11438, "Python id": 11442, "finished": true, "Ev Idx": 14342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915180.173, "dur": 6.506,
    "args": {
      "Python parent id": 11438, "Python id": 11443, "finished": true, "Ev Idx": 14343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915180.493, "dur": 6.105,
    "args": {
      "Python parent id": 11443, "Python id": 11444, "finished": true, "Ev Idx": 14344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915180.644, "dur": 3.957,
    "args": {
      "Python parent id": 11444, "Python id": 11445, "finished": true, "Ev Idx": 14345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915180.884, "dur": 3.641,
    "args": {
      "Python parent id": 11445, "Python id": 11446, "finished": true, "Ev Idx": 14346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915181.055, "dur": 0.304,
    "args": {
      "Python parent id": 11446, "Python id": 11447, "finished": true, "Ev Idx": 14347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915181.211, "dur": 0.057,
    "args": {
      "Python parent id": 11447, "Python id": 11448, "finished": true, "Ev Idx": 14348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915181.559, "dur": 2.529,
    "args": {
      "Python parent id": 11446, "Python id": 11449, "finished": true, "Ev Idx": 14349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915181.803, "dur": 2.133,
    "args": {
      "Python parent id": 11449, "Python id": 11450, "finished": true, "Ev Idx": 14350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915182.023, "dur": 1.853,
    "args": {
      "Python parent id": 11450, "Python id": 11451, "finished": true, "Ev Idx": 14351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915182.219, "dur": 1.448,
    "args": {
      "Python parent id": 11451, "Python id": 11452, "finished": true, "Ev Idx": 14352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915182.447, "dur": 0.590,
    "args": {
      "Python parent id": 11452, "Python id": 11453, "finished": true, "Ev Idx": 14353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915182.548, "dur": 0.072,
    "args": {
      "Python parent id": 11453, "Python id": 11454, "finished": true, "Ev Idx": 14354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915182.817, "dur": 0.177,
    "args": {
      "Python parent id": 11453, "Python id": 11455, "finished": true, "Ev Idx": 14355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915184.323, "dur": 0.163,
    "args": {
      "Python parent id": 11446, "Python id": 11456, "finished": true, "Ev Idx": 14356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915184.825, "dur": 0.116,
    "args": {
      "Python parent id": 11444, "Python id": 11457, "finished": true, "Ev Idx": 14357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915185.223, "dur": 1.310,
    "args": {
      "Python parent id": 11444, "Python id": 11458, "finished": true, "Ev Idx": 14358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915185.419, "dur": 1.070,
    "args": {
      "Python parent id": 11458, "Python id": 11459, "finished": true, "Ev Idx": 14359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915185.565, "dur": 0.486,
    "args": {
      "Python parent id": 11459, "Python id": 11460, "finished": true, "Ev Idx": 14360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915185.771, "dur": 0.222,
    "args": {
      "Python parent id": 11460, "Python id": 11461, "finished": true, "Ev Idx": 14361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915185.878, "dur": 0.082,
    "args": {
      "Python parent id": 11461, "Python id": 11462, "finished": true, "Ev Idx": 14362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915186.177, "dur": 0.266,
    "args": {
      "Python parent id": 11459, "Python id": 11463, "finished": true, "Ev Idx": 14363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915187.026, "dur": 0.257,
    "args": {
      "Python parent id": 11428, "Python id": 11464, "finished": true, "Ev Idx": 14364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915188.042, "dur": 1.255,
    "args": {
      "Python parent id": 11428, "Python id": 11465, "finished": true, "Ev Idx": 14365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915188.618, "dur": 0.630,
    "args": {
      "Python parent id": 11465, "Python id": 11466, "finished": true, "Ev Idx": 14366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915189.791, "dur": 3.370,
    "args": {
      "Python parent id": 11427, "Python id": 11467, "finished": true, "Ev Idx": 14367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915193.797, "dur": 8.250,
    "args": {
      "Python parent id": 12, "Python id": 11468, "finished": true, "Ev Idx": 14368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915202.594, "dur": 0.111,
    "args": {
      "Python parent id": 12, "Python id": 11469, "finished": true, "Ev Idx": 14369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915203.402, "dur": 0.184,
    "args": {
      "Python parent id": 12, "Python id": 11470, "finished": true, "Ev Idx": 14370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915204.109, "dur": 0.081,
    "args": {
      "Python parent id": 12, "Python id": 11471, "finished": true, "Ev Idx": 14371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915204.596, "dur": 10.275,
    "args": {
      "Python parent id": 12, "Python id": 11472, "finished": true, "Ev Idx": 14372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915216.105, "dur": 15.905,
    "args": {
      "Python parent id": 12, "Python id": 11473, "finished": true, "Ev Idx": 14373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915216.453, "dur": 11.821,
    "args": {
      "Python parent id": 11473, "Python id": 11474, "finished": true, "Ev Idx": 14374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915216.622, "dur": 0.533,
    "args": {
      "Python parent id": 11474, "Python id": 11475, "finished": true, "Ev Idx": 14375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915216.775, "dur": 0.277,
    "args": {
      "Python parent id": 11475, "Python id": 11476, "finished": true, "Ev Idx": 14376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915216.873, "dur": 0.112,
    "args": {
      "Python parent id": 11476, "Python id": 11477, "finished": true, "Ev Idx": 14377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915217.508, "dur": 8.712,
    "args": {
      "Python parent id": 11474, "Python id": 11478, "finished": true, "Ev Idx": 14378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915217.614, "dur": 0.117,
    "args": {
      "Python parent id": 11478, "Python id": 11479, "finished": true, "Ev Idx": 14379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915217.836, "dur": 0.070,
    "args": {
      "Python parent id": 11478, "Python id": 11480, "finished": true, "Ev Idx": 14380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915218.060, "dur": 0.057,
    "args": {
      "Python parent id": 11478, "Python id": 11481, "finished": true, "Ev Idx": 14381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915218.521, "dur": 0.039,
    "args": {
      "Python parent id": 11478, "Python id": 11482, "finished": true, "Ev Idx": 14382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915218.837, "dur": 0.080,
    "args": {
      "Python parent id": 11478, "Python id": 11483, "finished": true, "Ev Idx": 14383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915219.285, "dur": 6.862,
    "args": {
      "Python parent id": 11478, "Python id": 11484, "finished": true, "Ev Idx": 14384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915219.384, "dur": 0.065,
    "args": {
      "Python parent id": 11484, "Python id": 11485, "finished": true, "Ev Idx": 14385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915219.607, "dur": 0.061,
    "args": {
      "Python parent id": 11484, "Python id": 11486, "finished": true, "Ev Idx": 14386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915219.877, "dur": 0.074,
    "args": {
      "Python parent id": 11484, "Python id": 11487, "finished": true, "Ev Idx": 14387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915220.293, "dur": 0.044,
    "args": {
      "Python parent id": 11484, "Python id": 11488, "finished": true, "Ev Idx": 14388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915220.663, "dur": 5.403,
    "args": {
      "Python parent id": 11484, "Python id": 11489, "finished": true, "Ev Idx": 14389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915220.889, "dur": 5.102,
    "args": {
      "Python parent id": 11489, "Python id": 11490, "finished": true, "Ev Idx": 14390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915221.032, "dur": 3.220,
    "args": {
      "Python parent id": 11490, "Python id": 11491, "finished": true, "Ev Idx": 14391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915221.225, "dur": 2.981,
    "args": {
      "Python parent id": 11491, "Python id": 11492, "finished": true, "Ev Idx": 14392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915221.360, "dur": 0.269,
    "args": {
      "Python parent id": 11492, "Python id": 11493, "finished": true, "Ev Idx": 14393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915221.484, "dur": 0.066,
    "args": {
      "Python parent id": 11493, "Python id": 11494, "finished": true, "Ev Idx": 14394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915221.835, "dur": 2.052,
    "args": {
      "Python parent id": 11492, "Python id": 11495, "finished": true, "Ev Idx": 14395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.054, "dur": 1.733,
    "args": {
      "Python parent id": 11495, "Python id": 11496, "finished": true, "Ev Idx": 14396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.225, "dur": 1.513,
    "args": {
      "Python parent id": 11496, "Python id": 11497, "finished": true, "Ev Idx": 14397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.350, "dur": 1.162,
    "args": {
      "Python parent id": 11497, "Python id": 11498, "finished": true, "Ev Idx": 14398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.505, "dur": 0.512,
    "args": {
      "Python parent id": 11498, "Python id": 11499, "finished": true, "Ev Idx": 14399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.601, "dur": 0.056,
    "args": {
      "Python parent id": 11499, "Python id": 11500, "finished": true, "Ev Idx": 14400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915222.820, "dur": 0.141,
    "args": {
      "Python parent id": 11499, "Python id": 11501, "finished": true, "Ev Idx": 14401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915224.079, "dur": 0.075,
    "args": {
      "Python parent id": 11492, "Python id": 11502, "finished": true, "Ev Idx": 14402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915224.425, "dur": 0.075,
    "args": {
      "Python parent id": 11490, "Python id": 11503, "finished": true, "Ev Idx": 14403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915224.744, "dur": 1.183,
    "args": {
      "Python parent id": 11490, "Python id": 11504, "finished": true, "Ev Idx": 14404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915224.923, "dur": 0.962,
    "args": {
      "Python parent id": 11504, "Python id": 11505, "finished": true, "Ev Idx": 14405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915225.062, "dur": 0.410,
    "args": {
      "Python parent id": 11505, "Python id": 11506, "finished": true, "Ev Idx": 14406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915225.201, "dur": 0.211,
    "args": {
      "Python parent id": 11506, "Python id": 11507, "finished": true, "Ev Idx": 14407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915225.305, "dur": 0.058,
    "args": {
      "Python parent id": 11507, "Python id": 11508, "finished": true, "Ev Idx": 14408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915225.616, "dur": 0.236,
    "args": {
      "Python parent id": 11505, "Python id": 11509, "finished": true, "Ev Idx": 14409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915226.348, "dur": 0.172,
    "args": {
      "Python parent id": 11474, "Python id": 11510, "finished": true, "Ev Idx": 14410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915227.190, "dur": 0.949,
    "args": {
      "Python parent id": 11474, "Python id": 11511, "finished": true, "Ev Idx": 14411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915227.637, "dur": 0.449,
    "args": {
      "Python parent id": 11511, "Python id": 11512, "finished": true, "Ev Idx": 14412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915228.559, "dur": 3.388,
    "args": {
      "Python parent id": 11473, "Python id": 11513, "finished": true, "Ev Idx": 14413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915232.949, "dur": 119.377,
    "args": {
      "Python parent id": 12, "Python id": 11514, "finished": true, "Ev Idx": 14414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915233.354, "dur": 0.684,
    "args": {
      "Python parent id": 11514, "Python id": 11515, "finished": true, "Ev Idx": 14415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915233.628, "dur": 0.343,
    "args": {
      "Python parent id": 11515, "Python id": 11516, "finished": true, "Ev Idx": 14416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915234.386, "dur": 0.408,
    "args": {
      "Python parent id": 11514, "Python id": 11517, "finished": true, "Ev Idx": 14417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915235.006, "dur": 0.265,
    "args": {
      "Python parent id": 11514, "Python id": 11518, "finished": true, "Ev Idx": 14418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915235.518, "dur": 1.603,
    "args": {
      "Python parent id": 11514, "Python id": 11519, "finished": true, "Ev Idx": 14419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915237.745, "dur": 114.468,
    "args": {
      "Python parent id": 11514, "Python id": 11520, "finished": true, "Ev Idx": 14420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915238.741, "dur": 113.210,
    "args": {
      "Python parent id": 11520, "Python id": 11521, "finished": true, "Ev Idx": 14421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915239.069, "dur": 112.781,
    "args": {
      "Python parent id": 11521, "Python id": 11522, "finished": true, "Ev Idx": 14422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915244.730, "dur": 105.335,
    "args": {
      "Python parent id": 11522, "Python id": 11523, "finished": true, "Ev Idx": 14423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915245.101, "dur": 104.905,
    "args": {
      "Python parent id": 11523, "Python id": 11524, "finished": true, "Ev Idx": 14424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915245.667, "dur": 104.173,
    "args": {
      "Python parent id": 11524, "Python id": 11525, "finished": true, "Ev Idx": 14425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915246.232, "dur": 0.531,
    "args": {
      "Python parent id": 11525, "Python id": 11526, "finished": true, "Ev Idx": 14426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915246.532, "dur": 0.191,
    "args": {
      "Python parent id": 11526, "Python id": 11527, "finished": true, "Ev Idx": 14427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915247.093, "dur": 0.408,
    "args": {
      "Python parent id": 11525, "Python id": 11528, "finished": true, "Ev Idx": 14428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915247.359, "dur": 0.094,
    "args": {
      "Python parent id": 11528, "Python id": 11529, "finished": true, "Ev Idx": 14429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915247.864, "dur": 101.874,
    "args": {
      "Python parent id": 11525, "Python id": 11530, "finished": true, "Ev Idx": 14430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915248.546, "dur": 99.477,
    "args": {
      "Python parent id": 11530, "Python id": 11531, "finished": true, "Ev Idx": 14431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915255.947, "dur": 20.521,
    "args": {
      "Python parent id": 11531, "Python id": 11532, "finished": true, "Ev Idx": 14432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915256.097, "dur": 0.125,
    "args": {
      "Python parent id": 11532, "Python id": 11533, "finished": true, "Ev Idx": 14433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915256.574, "dur": 19.743,
    "args": {
      "Python parent id": 11532, "Python id": 11534, "finished": true, "Ev Idx": 14434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915256.698, "dur": 7.536,
    "args": {
      "Python parent id": 11534, "Python id": 11535, "finished": true, "Ev Idx": 14435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915296.282, "dur": 11.958,
    "args": {
      "Python parent id": 11531, "Python id": 11536, "finished": true, "Ev Idx": 14436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915320.193, "dur": 10.246,
    "args": {
      "Python parent id": 11531, "Python id": 11537, "finished": true, "Ev Idx": 14437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915353.922, "dur": 20.591,
    "args": {
      "Python parent id": 12, "Python id": 11538, "finished": true, "Ev Idx": 14438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915354.396, "dur": 15.181,
    "args": {
      "Python parent id": 11538, "Python id": 11539, "finished": true, "Ev Idx": 14439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915354.619, "dur": 0.687,
    "args": {
      "Python parent id": 11539, "Python id": 11540, "finished": true, "Ev Idx": 14440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915354.805, "dur": 0.402,
    "args": {
      "Python parent id": 11540, "Python id": 11541, "finished": true, "Ev Idx": 14441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915355.004, "dur": 0.149,
    "args": {
      "Python parent id": 11541, "Python id": 11542, "finished": true, "Ev Idx": 14442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915355.786, "dur": 10.731,
    "args": {
      "Python parent id": 11539, "Python id": 11543, "finished": true, "Ev Idx": 14443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915355.918, "dur": 0.105,
    "args": {
      "Python parent id": 11543, "Python id": 11544, "finished": true, "Ev Idx": 14444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915356.180, "dur": 0.063,
    "args": {
      "Python parent id": 11543, "Python id": 11545, "finished": true, "Ev Idx": 14445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915356.399, "dur": 0.061,
    "args": {
      "Python parent id": 11543, "Python id": 11546, "finished": true, "Ev Idx": 14446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915356.971, "dur": 0.030,
    "args": {
      "Python parent id": 11543, "Python id": 11547, "finished": true, "Ev Idx": 14447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915357.328, "dur": 0.111,
    "args": {
      "Python parent id": 11543, "Python id": 11548, "finished": true, "Ev Idx": 14448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915357.822, "dur": 8.642,
    "args": {
      "Python parent id": 11543, "Python id": 11549, "finished": true, "Ev Idx": 14449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915357.927, "dur": 0.062,
    "args": {
      "Python parent id": 11549, "Python id": 11550, "finished": true, "Ev Idx": 14450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915358.190, "dur": 0.060,
    "args": {
      "Python parent id": 11549, "Python id": 11551, "finished": true, "Ev Idx": 14451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915358.457, "dur": 0.062,
    "args": {
      "Python parent id": 11549, "Python id": 11552, "finished": true, "Ev Idx": 14452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915358.896, "dur": 0.049,
    "args": {
      "Python parent id": 11549, "Python id": 11553, "finished": true, "Ev Idx": 14453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915359.271, "dur": 7.064,
    "args": {
      "Python parent id": 11549, "Python id": 11554, "finished": true, "Ev Idx": 14454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915359.533, "dur": 6.716,
    "args": {
      "Python parent id": 11554, "Python id": 11555, "finished": true, "Ev Idx": 14455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915359.674, "dur": 4.476,
    "args": {
      "Python parent id": 11555, "Python id": 11556, "finished": true, "Ev Idx": 14456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915359.869, "dur": 4.202,
    "args": {
      "Python parent id": 11556, "Python id": 11557, "finished": true, "Ev Idx": 14457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915359.999, "dur": 0.357,
    "args": {
      "Python parent id": 11557, "Python id": 11558, "finished": true, "Ev Idx": 14458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915360.198, "dur": 0.086,
    "args": {
      "Python parent id": 11558, "Python id": 11559, "finished": true, "Ev Idx": 14459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915360.582, "dur": 3.061,
    "args": {
      "Python parent id": 11557, "Python id": 11560, "finished": true, "Ev Idx": 14460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915360.857, "dur": 2.704,
    "args": {
      "Python parent id": 11560, "Python id": 11561, "finished": true, "Ev Idx": 14461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915361.130, "dur": 2.384,
    "args": {
      "Python parent id": 11561, "Python id": 11562, "finished": true, "Ev Idx": 14462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915361.333, "dur": 1.820,
    "args": {
      "Python parent id": 11562, "Python id": 11563, "finished": true, "Ev Idx": 14463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915361.612, "dur": 0.705,
    "args": {
      "Python parent id": 11563, "Python id": 11564, "finished": true, "Ev Idx": 14464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915361.736, "dur": 0.066,
    "args": {
      "Python parent id": 11564, "Python id": 11565, "finished": true, "Ev Idx": 14465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915362.046, "dur": 0.199,
    "args": {
      "Python parent id": 11564, "Python id": 11566, "finished": true, "Ev Idx": 14466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915363.836, "dur": 0.190,
    "args": {
      "Python parent id": 11557, "Python id": 11567, "finished": true, "Ev Idx": 14467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915364.394, "dur": 0.107,
    "args": {
      "Python parent id": 11555, "Python id": 11568, "finished": true, "Ev Idx": 14468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915364.783, "dur": 1.404,
    "args": {
      "Python parent id": 11555, "Python id": 11569, "finished": true, "Ev Idx": 14469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915364.977, "dur": 1.164,
    "args": {
      "Python parent id": 11569, "Python id": 11570, "finished": true, "Ev Idx": 14470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915365.110, "dur": 0.417,
    "args": {
      "Python parent id": 11570, "Python id": 11571, "finished": true, "Ev Idx": 14471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915365.274, "dur": 0.195,
    "args": {
      "Python parent id": 11571, "Python id": 11572, "finished": true, "Ev Idx": 14472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915365.380, "dur": 0.058,
    "args": {
      "Python parent id": 11572, "Python id": 11573, "finished": true, "Ev Idx": 14473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915365.701, "dur": 0.416,
    "args": {
      "Python parent id": 11570, "Python id": 11574, "finished": true, "Ev Idx": 14474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915366.694, "dur": 0.328,
    "args": {
      "Python parent id": 11539, "Python id": 11575, "finished": true, "Ev Idx": 14475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915367.810, "dur": 1.583,
    "args": {
      "Python parent id": 11539, "Python id": 11576, "finished": true, "Ev Idx": 14476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915368.514, "dur": 0.788,
    "args": {
      "Python parent id": 11576, "Python id": 11577, "finished": true, "Ev Idx": 14477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915370.046, "dur": 4.375,
    "args": {
      "Python parent id": 11538, "Python id": 11578, "finished": true, "Ev Idx": 14478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915375.159, "dur": 8.187,
    "args": {
      "Python parent id": 12, "Python id": 11579, "finished": true, "Ev Idx": 14479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915383.884, "dur": 0.132,
    "args": {
      "Python parent id": 12, "Python id": 11580, "finished": true, "Ev Idx": 14480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915384.836, "dur": 0.226,
    "args": {
      "Python parent id": 12, "Python id": 11581, "finished": true, "Ev Idx": 14481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915385.582, "dur": 0.077,
    "args": {
      "Python parent id": 12, "Python id": 11582, "finished": true, "Ev Idx": 14482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915386.068, "dur": 10.918,
    "args": {
      "Python parent id": 12, "Python id": 11583, "finished": true, "Ev Idx": 14483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915398.467, "dur": 17.332,
    "args": {
      "Python parent id": 12, "Python id": 11584, "finished": true, "Ev Idx": 14484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915398.803, "dur": 12.605,
    "args": {
      "Python parent id": 11584, "Python id": 11585, "finished": true, "Ev Idx": 14485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915398.995, "dur": 0.576,
    "args": {
      "Python parent id": 11585, "Python id": 11586, "finished": true, "Ev Idx": 14486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915399.163, "dur": 0.266,
    "args": {
      "Python parent id": 11586, "Python id": 11587, "finished": true, "Ev Idx": 14487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915399.300, "dur": 0.093,
    "args": {
      "Python parent id": 11587, "Python id": 11588, "finished": true, "Ev Idx": 14488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915399.914, "dur": 9.338,
    "args": {
      "Python parent id": 11585, "Python id": 11589, "finished": true, "Ev Idx": 14489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915399.995, "dur": 0.068,
    "args": {
      "Python parent id": 11589, "Python id": 11590, "finished": true, "Ev Idx": 14490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915400.213, "dur": 0.060,
    "args": {
      "Python parent id": 11589, "Python id": 11591, "finished": true, "Ev Idx": 14491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915400.420, "dur": 0.059,
    "args": {
      "Python parent id": 11589, "Python id": 11592, "finished": true, "Ev Idx": 14492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915400.939, "dur": 0.038,
    "args": {
      "Python parent id": 11589, "Python id": 11593, "finished": true, "Ev Idx": 14493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915401.274, "dur": 0.069,
    "args": {
      "Python parent id": 11589, "Python id": 11594, "finished": true, "Ev Idx": 14494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915401.701, "dur": 7.499,
    "args": {
      "Python parent id": 11589, "Python id": 11595, "finished": true, "Ev Idx": 14495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915401.801, "dur": 0.056,
    "args": {
      "Python parent id": 11595, "Python id": 11596, "finished": true, "Ev Idx": 14496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915402.006, "dur": 0.056,
    "args": {
      "Python parent id": 11595, "Python id": 11597, "finished": true, "Ev Idx": 14497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915402.239, "dur": 0.065,
    "args": {
      "Python parent id": 11595, "Python id": 11598, "finished": true, "Ev Idx": 14498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915402.669, "dur": 0.036,
    "args": {
      "Python parent id": 11595, "Python id": 11599, "finished": true, "Ev Idx": 14499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.094, "dur": 5.978,
    "args": {
      "Python parent id": 11595, "Python id": 11600, "finished": true, "Ev Idx": 14500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.319, "dur": 5.673,
    "args": {
      "Python parent id": 11600, "Python id": 11601, "finished": true, "Ev Idx": 14501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.494, "dur": 3.535,
    "args": {
      "Python parent id": 11601, "Python id": 11602, "finished": true, "Ev Idx": 14502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.659, "dur": 3.302,
    "args": {
      "Python parent id": 11602, "Python id": 11603, "finished": true, "Ev Idx": 14503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.786, "dur": 0.327,
    "args": {
      "Python parent id": 11603, "Python id": 11604, "finished": true, "Ev Idx": 14504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915403.962, "dur": 0.080,
    "args": {
      "Python parent id": 11604, "Python id": 11605, "finished": true, "Ev Idx": 14505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915404.283, "dur": 2.362,
    "args": {
      "Python parent id": 11603, "Python id": 11606, "finished": true, "Ev Idx": 14506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915404.505, "dur": 2.082,
    "args": {
      "Python parent id": 11606, "Python id": 11607, "finished": true, "Ev Idx": 14507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915404.762, "dur": 1.757,
    "args": {
      "Python parent id": 11607, "Python id": 11608, "finished": true, "Ev Idx": 14508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915404.916, "dur": 1.344,
    "args": {
      "Python parent id": 11608, "Python id": 11609, "finished": true, "Ev Idx": 14509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915405.123, "dur": 0.548,
    "args": {
      "Python parent id": 11609, "Python id": 11610, "finished": true, "Ev Idx": 14510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915405.225, "dur": 0.042,
    "args": {
      "Python parent id": 11610, "Python id": 11611, "finished": true, "Ev Idx": 14511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915405.470, "dur": 0.155,
    "args": {
      "Python parent id": 11610, "Python id": 11612, "finished": true, "Ev Idx": 14512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915406.850, "dur": 0.072,
    "args": {
      "Python parent id": 11603, "Python id": 11613, "finished": true, "Ev Idx": 14513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915407.242, "dur": 0.112,
    "args": {
      "Python parent id": 11601, "Python id": 11614, "finished": true, "Ev Idx": 14514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915407.657, "dur": 1.264,
    "args": {
      "Python parent id": 11601, "Python id": 11615, "finished": true, "Ev Idx": 14515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915407.849, "dur": 1.026,
    "args": {
      "Python parent id": 11615, "Python id": 11616, "finished": true, "Ev Idx": 14516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915407.996, "dur": 0.423,
    "args": {
      "Python parent id": 11616, "Python id": 11617, "finished": true, "Ev Idx": 14517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915408.151, "dur": 0.208,
    "args": {
      "Python parent id": 11617, "Python id": 11618, "finished": true, "Ev Idx": 14518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915408.261, "dur": 0.059,
    "args": {
      "Python parent id": 11618, "Python id": 11619, "finished": true, "Ev Idx": 14519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915408.598, "dur": 0.254,
    "args": {
      "Python parent id": 11616, "Python id": 11620, "finished": true, "Ev Idx": 14520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915409.380, "dur": 0.209,
    "args": {
      "Python parent id": 11585, "Python id": 11621, "finished": true, "Ev Idx": 14521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915410.189, "dur": 1.081,
    "args": {
      "Python parent id": 11585, "Python id": 11622, "finished": true, "Ev Idx": 14522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915410.710, "dur": 0.511,
    "args": {
      "Python parent id": 11622, "Python id": 11623, "finished": true, "Ev Idx": 14523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915411.777, "dur": 3.962,
    "args": {
      "Python parent id": 11584, "Python id": 11624, "finished": true, "Ev Idx": 14524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915416.870, "dur": 97.521,
    "args": {
      "Python parent id": 12, "Python id": 11625, "finished": true, "Ev Idx": 14525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915417.254, "dur": 0.486,
    "args": {
      "Python parent id": 11625, "Python id": 11626, "finished": true, "Ev Idx": 14526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915417.603, "dur": 0.072,
    "args": {
      "Python parent id": 11626, "Python id": 11627, "finished": true, "Ev Idx": 14527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915418.132, "dur": 0.522,
    "args": {
      "Python parent id": 11625, "Python id": 11628, "finished": true, "Ev Idx": 14528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915418.888, "dur": 0.246,
    "args": {
      "Python parent id": 11625, "Python id": 11629, "finished": true, "Ev Idx": 14529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915419.423, "dur": 2.233,
    "args": {
      "Python parent id": 11625, "Python id": 11630, "finished": true, "Ev Idx": 14530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915422.340, "dur": 91.947,
    "args": {
      "Python parent id": 11625, "Python id": 11631, "finished": true, "Ev Idx": 14531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915423.371, "dur": 90.688,
    "args": {
      "Python parent id": 11631, "Python id": 11632, "finished": true, "Ev Idx": 14532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915423.761, "dur": 90.237,
    "args": {
      "Python parent id": 11632, "Python id": 11633, "finished": true, "Ev Idx": 14533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915430.335, "dur": 82.140,
    "args": {
      "Python parent id": 11633, "Python id": 11634, "finished": true, "Ev Idx": 14534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915430.669, "dur": 81.702,
    "args": {
      "Python parent id": 11634, "Python id": 11635, "finished": true, "Ev Idx": 14535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915431.270, "dur": 80.923,
    "args": {
      "Python parent id": 11635, "Python id": 11636, "finished": true, "Ev Idx": 14536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915431.877, "dur": 0.488,
    "args": {
      "Python parent id": 11636, "Python id": 11637, "finished": true, "Ev Idx": 14537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915432.178, "dur": 0.121,
    "args": {
      "Python parent id": 11637, "Python id": 11638, "finished": true, "Ev Idx": 14538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915432.730, "dur": 0.366,
    "args": {
      "Python parent id": 11636, "Python id": 11639, "finished": true, "Ev Idx": 14539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915432.982, "dur": 0.067,
    "args": {
      "Python parent id": 11639, "Python id": 11640, "finished": true, "Ev Idx": 14540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915433.439, "dur": 78.640,
    "args": {
      "Python parent id": 11636, "Python id": 11641, "finished": true, "Ev Idx": 14541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915434.208, "dur": 76.421,
    "args": {
      "Python parent id": 11641, "Python id": 11642, "finished": true, "Ev Idx": 14542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915442.297, "dur": 20.725,
    "args": {
      "Python parent id": 11642, "Python id": 11643, "finished": true, "Ev Idx": 14543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915442.447, "dur": 0.096,
    "args": {
      "Python parent id": 11643, "Python id": 11644, "finished": true, "Ev Idx": 14544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915442.944, "dur": 19.940,
    "args": {
      "Python parent id": 11643, "Python id": 11645, "finished": true, "Ev Idx": 14545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915443.091, "dur": 9.611,
    "args": {
      "Python parent id": 11645, "Python id": 11646, "finished": true, "Ev Idx": 14546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915470.925, "dur": 7.357,
    "args": {
      "Python parent id": 11642, "Python id": 11647, "finished": true, "Ev Idx": 14547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915486.796, "dur": 8.213,
    "args": {
      "Python parent id": 11642, "Python id": 11648, "finished": true, "Ev Idx": 14548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915515.733, "dur": 18.893,
    "args": {
      "Python parent id": 12, "Python id": 11649, "finished": true, "Ev Idx": 14549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915516.200, "dur": 14.284,
    "args": {
      "Python parent id": 11649, "Python id": 11650, "finished": true, "Ev Idx": 14550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915516.388, "dur": 0.580,
    "args": {
      "Python parent id": 11650, "Python id": 11651, "finished": true, "Ev Idx": 14551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915516.554, "dur": 0.306,
    "args": {
      "Python parent id": 11651, "Python id": 11652, "finished": true, "Ev Idx": 14552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915516.696, "dur": 0.129,
    "args": {
      "Python parent id": 11652, "Python id": 11653, "finished": true, "Ev Idx": 14553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915517.318, "dur": 10.435,
    "args": {
      "Python parent id": 11650, "Python id": 11654, "finished": true, "Ev Idx": 14554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915517.466, "dur": 0.069,
    "args": {
      "Python parent id": 11654, "Python id": 11655, "finished": true, "Ev Idx": 14555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915517.706, "dur": 0.062,
    "args": {
      "Python parent id": 11654, "Python id": 11656, "finished": true, "Ev Idx": 14556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915517.917, "dur": 0.061,
    "args": {
      "Python parent id": 11654, "Python id": 11657, "finished": true, "Ev Idx": 14557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915518.500, "dur": 0.036,
    "args": {
      "Python parent id": 11654, "Python id": 11658, "finished": true, "Ev Idx": 14558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915518.842, "dur": 0.111,
    "args": {
      "Python parent id": 11654, "Python id": 11659, "finished": true, "Ev Idx": 14559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915519.310, "dur": 8.366,
    "args": {
      "Python parent id": 11654, "Python id": 11660, "finished": true, "Ev Idx": 14560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915519.418, "dur": 0.061,
    "args": {
      "Python parent id": 11660, "Python id": 11661, "finished": true, "Ev Idx": 14561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915519.645, "dur": 0.056,
    "args": {
      "Python parent id": 11660, "Python id": 11662, "finished": true, "Ev Idx": 14562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915519.939, "dur": 0.073,
    "args": {
      "Python parent id": 11660, "Python id": 11663, "finished": true, "Ev Idx": 14563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915520.436, "dur": 0.036,
    "args": {
      "Python parent id": 11660, "Python id": 11664, "finished": true, "Ev Idx": 14564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915520.816, "dur": 6.780,
    "args": {
      "Python parent id": 11660, "Python id": 11665, "finished": true, "Ev Idx": 14565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915521.097, "dur": 6.419,
    "args": {
      "Python parent id": 11665, "Python id": 11666, "finished": true, "Ev Idx": 14566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915521.247, "dur": 4.145,
    "args": {
      "Python parent id": 11666, "Python id": 11667, "finished": true, "Ev Idx": 14567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915521.462, "dur": 3.885,
    "args": {
      "Python parent id": 11667, "Python id": 11668, "finished": true, "Ev Idx": 14568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915521.616, "dur": 0.334,
    "args": {
      "Python parent id": 11668, "Python id": 11669, "finished": true, "Ev Idx": 14569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915521.813, "dur": 0.063,
    "args": {
      "Python parent id": 11669, "Python id": 11670, "finished": true, "Ev Idx": 14570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915522.160, "dur": 2.764,
    "args": {
      "Python parent id": 11668, "Python id": 11671, "finished": true, "Ev Idx": 14571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915522.437, "dur": 2.363,
    "args": {
      "Python parent id": 11671, "Python id": 11672, "finished": true, "Ev Idx": 14572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915522.708, "dur": 2.032,
    "args": {
      "Python parent id": 11672, "Python id": 11673, "finished": true, "Ev Idx": 14573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915522.932, "dur": 1.488,
    "args": {
      "Python parent id": 11673, "Python id": 11674, "finished": true, "Ev Idx": 14574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915523.185, "dur": 0.591,
    "args": {
      "Python parent id": 11674, "Python id": 11675, "finished": true, "Ev Idx": 14575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915523.273, "dur": 0.068,
    "args": {
      "Python parent id": 11675, "Python id": 11676, "finished": true, "Ev Idx": 14576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915523.522, "dur": 0.189,
    "args": {
      "Python parent id": 11675, "Python id": 11677, "finished": true, "Ev Idx": 14577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915525.150, "dur": 0.163,
    "args": {
      "Python parent id": 11668, "Python id": 11678, "finished": true, "Ev Idx": 14578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915525.631, "dur": 0.117,
    "args": {
      "Python parent id": 11666, "Python id": 11679, "finished": true, "Ev Idx": 14579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915526.073, "dur": 1.373,
    "args": {
      "Python parent id": 11666, "Python id": 11680, "finished": true, "Ev Idx": 14580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915526.277, "dur": 1.122,
    "args": {
      "Python parent id": 11680, "Python id": 11681, "finished": true, "Ev Idx": 14581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915526.409, "dur": 0.468,
    "args": {
      "Python parent id": 11681, "Python id": 11682, "finished": true, "Ev Idx": 14582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915526.549, "dur": 0.251,
    "args": {
      "Python parent id": 11682, "Python id": 11683, "finished": true, "Ev Idx": 14583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915526.663, "dur": 0.082,
    "args": {
      "Python parent id": 11683, "Python id": 11684, "finished": true, "Ev Idx": 14584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915527.020, "dur": 0.337,
    "args": {
      "Python parent id": 11681, "Python id": 11685, "finished": true, "Ev Idx": 14585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915527.928, "dur": 0.289,
    "args": {
      "Python parent id": 11650, "Python id": 11686, "finished": true, "Ev Idx": 14586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915528.969, "dur": 1.381,
    "args": {
      "Python parent id": 11650, "Python id": 11687, "finished": true, "Ev Idx": 14587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915529.663, "dur": 0.611,
    "args": {
      "Python parent id": 11687, "Python id": 11688, "finished": true, "Ev Idx": 14588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915530.877, "dur": 3.667,
    "args": {
      "Python parent id": 11649, "Python id": 11689, "finished": true, "Ev Idx": 14589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915535.313, "dur": 8.372,
    "args": {
      "Python parent id": 12, "Python id": 11690, "finished": true, "Ev Idx": 14590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915544.203, "dur": 0.128,
    "args": {
      "Python parent id": 12, "Python id": 11691, "finished": true, "Ev Idx": 14591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915545.023, "dur": 0.209,
    "args": {
      "Python parent id": 12, "Python id": 11692, "finished": true, "Ev Idx": 14592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915545.755, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 11693, "finished": true, "Ev Idx": 14593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915546.254, "dur": 10.374,
    "args": {
      "Python parent id": 12, "Python id": 11694, "finished": true, "Ev Idx": 14594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915557.989, "dur": 16.639,
    "args": {
      "Python parent id": 12, "Python id": 11695, "finished": true, "Ev Idx": 14595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915558.332, "dur": 12.166,
    "args": {
      "Python parent id": 11695, "Python id": 11696, "finished": true, "Ev Idx": 14596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915558.483, "dur": 0.502,
    "args": {
      "Python parent id": 11696, "Python id": 11697, "finished": true, "Ev Idx": 14597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915558.636, "dur": 0.281,
    "args": {
      "Python parent id": 11697, "Python id": 11698, "finished": true, "Ev Idx": 14598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915558.779, "dur": 0.103,
    "args": {
      "Python parent id": 11698, "Python id": 11699, "finished": true, "Ev Idx": 14599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915559.226, "dur": 9.220,
    "args": {
      "Python parent id": 11696, "Python id": 11700, "finished": true, "Ev Idx": 14600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915559.365, "dur": 0.068,
    "args": {
      "Python parent id": 11700, "Python id": 11701, "finished": true, "Ev Idx": 14601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915559.540, "dur": 0.063,
    "args": {
      "Python parent id": 11700, "Python id": 11702, "finished": true, "Ev Idx": 14602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915559.745, "dur": 0.055,
    "args": {
      "Python parent id": 11700, "Python id": 11703, "finished": true, "Ev Idx": 14603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915560.238, "dur": 0.037,
    "args": {
      "Python parent id": 11700, "Python id": 11704, "finished": true, "Ev Idx": 14604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915560.603, "dur": 0.077,
    "args": {
      "Python parent id": 11700, "Python id": 11705, "finished": true, "Ev Idx": 14605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915561.092, "dur": 7.297,
    "args": {
      "Python parent id": 11700, "Python id": 11706, "finished": true, "Ev Idx": 14606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915561.187, "dur": 0.083,
    "args": {
      "Python parent id": 11706, "Python id": 11707, "finished": true, "Ev Idx": 14607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915561.451, "dur": 0.065,
    "args": {
      "Python parent id": 11706, "Python id": 11708, "finished": true, "Ev Idx": 14608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915561.699, "dur": 0.073,
    "args": {
      "Python parent id": 11706, "Python id": 11709, "finished": true, "Ev Idx": 14609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915562.177, "dur": 0.038,
    "args": {
      "Python parent id": 11706, "Python id": 11710, "finished": true, "Ev Idx": 14610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915562.573, "dur": 5.735,
    "args": {
      "Python parent id": 11706, "Python id": 11711, "finished": true, "Ev Idx": 14611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915562.809, "dur": 5.420,
    "args": {
      "Python parent id": 11711, "Python id": 11712, "finished": true, "Ev Idx": 14612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915562.998, "dur": 3.482,
    "args": {
      "Python parent id": 11712, "Python id": 11713, "finished": true, "Ev Idx": 14613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915563.209, "dur": 3.227,
    "args": {
      "Python parent id": 11713, "Python id": 11714, "finished": true, "Ev Idx": 14614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915563.334, "dur": 0.333,
    "args": {
      "Python parent id": 11714, "Python id": 11715, "finished": true, "Ev Idx": 14615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915563.508, "dur": 0.066,
    "args": {
      "Python parent id": 11715, "Python id": 11716, "finished": true, "Ev Idx": 14616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915563.837, "dur": 2.289,
    "args": {
      "Python parent id": 11714, "Python id": 11717, "finished": true, "Ev Idx": 14617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.087, "dur": 1.949,
    "args": {
      "Python parent id": 11717, "Python id": 11718, "finished": true, "Ev Idx": 14618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.330, "dur": 1.642,
    "args": {
      "Python parent id": 11718, "Python id": 11719, "finished": true, "Ev Idx": 14619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.449, "dur": 1.238,
    "args": {
      "Python parent id": 11719, "Python id": 11720, "finished": true, "Ev Idx": 14620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.611, "dur": 0.534,
    "args": {
      "Python parent id": 11720, "Python id": 11721, "finished": true, "Ev Idx": 14621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.695, "dur": 0.049,
    "args": {
      "Python parent id": 11721, "Python id": 11722, "finished": true, "Ev Idx": 14622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915564.945, "dur": 0.155,
    "args": {
      "Python parent id": 11721, "Python id": 11723, "finished": true, "Ev Idx": 14623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915566.303, "dur": 0.095,
    "args": {
      "Python parent id": 11714, "Python id": 11724, "finished": true, "Ev Idx": 14624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915566.661, "dur": 0.088,
    "args": {
      "Python parent id": 11712, "Python id": 11725, "finished": true, "Ev Idx": 14625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.046, "dur": 1.136,
    "args": {
      "Python parent id": 11712, "Python id": 11726, "finished": true, "Ev Idx": 14626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.192, "dur": 0.946,
    "args": {
      "Python parent id": 11726, "Python id": 11727, "finished": true, "Ev Idx": 14627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.341, "dur": 0.374,
    "args": {
      "Python parent id": 11727, "Python id": 11728, "finished": true, "Ev Idx": 14628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.487, "dur": 0.169,
    "args": {
      "Python parent id": 11728, "Python id": 11729, "finished": true, "Ev Idx": 14629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.576, "dur": 0.049,
    "args": {
      "Python parent id": 11729, "Python id": 11730, "finished": true, "Ev Idx": 14630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915567.858, "dur": 0.256,
    "args": {
      "Python parent id": 11727, "Python id": 11731, "finished": true, "Ev Idx": 14631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915568.565, "dur": 0.219,
    "args": {
      "Python parent id": 11696, "Python id": 11732, "finished": true, "Ev Idx": 14632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915569.372, "dur": 1.003,
    "args": {
      "Python parent id": 11696, "Python id": 11733, "finished": true, "Ev Idx": 14633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915569.832, "dur": 0.494,
    "args": {
      "Python parent id": 11733, "Python id": 11734, "finished": true, "Ev Idx": 14634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915570.746, "dur": 3.822,
    "args": {
      "Python parent id": 11695, "Python id": 11735, "finished": true, "Ev Idx": 14635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915575.604, "dur": 92.182,
    "args": {
      "Python parent id": 12, "Python id": 11736, "finished": true, "Ev Idx": 14636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915575.976, "dur": 0.391,
    "args": {
      "Python parent id": 11736, "Python id": 11737, "finished": true, "Ev Idx": 14637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915576.262, "dur": 0.049,
    "args": {
      "Python parent id": 11737, "Python id": 11738, "finished": true, "Ev Idx": 14638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915576.704, "dur": 0.377,
    "args": {
      "Python parent id": 11736, "Python id": 11739, "finished": true, "Ev Idx": 14639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915577.339, "dur": 0.241,
    "args": {
      "Python parent id": 11736, "Python id": 11740, "finished": true, "Ev Idx": 14640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915577.848, "dur": 1.814,
    "args": {
      "Python parent id": 11736, "Python id": 11741, "finished": true, "Ev Idx": 14641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915580.352, "dur": 87.309,
    "args": {
      "Python parent id": 11736, "Python id": 11742, "finished": true, "Ev Idx": 14642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915581.349, "dur": 86.087,
    "args": {
      "Python parent id": 11742, "Python id": 11743, "finished": true, "Ev Idx": 14643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915581.706, "dur": 85.671,
    "args": {
      "Python parent id": 11743, "Python id": 11744, "finished": true, "Ev Idx": 14644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915587.373, "dur": 78.696,
    "args": {
      "Python parent id": 11744, "Python id": 11745, "finished": true, "Ev Idx": 14645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915587.722, "dur": 78.231,
    "args": {
      "Python parent id": 11745, "Python id": 11746, "finished": true, "Ev Idx": 14646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915588.369, "dur": 77.516,
    "args": {
      "Python parent id": 11746, "Python id": 11747, "finished": true, "Ev Idx": 14647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915588.877, "dur": 0.481,
    "args": {
      "Python parent id": 11747, "Python id": 11748, "finished": true, "Ev Idx": 14648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915589.189, "dur": 0.124,
    "args": {
      "Python parent id": 11748, "Python id": 11749, "finished": true, "Ev Idx": 14649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915589.698, "dur": 0.382,
    "args": {
      "Python parent id": 11747, "Python id": 11750, "finished": true, "Ev Idx": 14650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915589.950, "dur": 0.082,
    "args": {
      "Python parent id": 11750, "Python id": 11751, "finished": true, "Ev Idx": 14651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915590.416, "dur": 75.354,
    "args": {
      "Python parent id": 11747, "Python id": 11752, "finished": true, "Ev Idx": 14652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915591.091, "dur": 73.179,
    "args": {
      "Python parent id": 11752, "Python id": 11753, "finished": true, "Ev Idx": 14653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915598.460, "dur": 19.052,
    "args": {
      "Python parent id": 11753, "Python id": 11754, "finished": true, "Ev Idx": 14654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915598.621, "dur": 0.104,
    "args": {
      "Python parent id": 11754, "Python id": 11755, "finished": true, "Ev Idx": 14655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915599.156, "dur": 18.204,
    "args": {
      "Python parent id": 11754, "Python id": 11756, "finished": true, "Ev Idx": 14656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915599.296, "dur": 8.434,
    "args": {
      "Python parent id": 11756, "Python id": 11757, "finished": true, "Ev Idx": 14657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915625.048, "dur": 6.870,
    "args": {
      "Python parent id": 11753, "Python id": 11758, "finished": true, "Ev Idx": 14658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915640.675, "dur": 8.826,
    "args": {
      "Python parent id": 11753, "Python id": 11759, "finished": true, "Ev Idx": 14659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915669.201, "dur": 18.453,
    "args": {
      "Python parent id": 12, "Python id": 11760, "finished": true, "Ev Idx": 14660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915669.607, "dur": 13.987,
    "args": {
      "Python parent id": 11760, "Python id": 11761, "finished": true, "Ev Idx": 14661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915669.767, "dur": 0.545,
    "args": {
      "Python parent id": 11761, "Python id": 11762, "finished": true, "Ev Idx": 14662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915669.934, "dur": 0.283,
    "args": {
      "Python parent id": 11762, "Python id": 11763, "finished": true, "Ev Idx": 14663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915670.093, "dur": 0.074,
    "args": {
      "Python parent id": 11763, "Python id": 11764, "finished": true, "Ev Idx": 14664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915670.735, "dur": 10.237,
    "args": {
      "Python parent id": 11761, "Python id": 11765, "finished": true, "Ev Idx": 14665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915670.838, "dur": 0.095,
    "args": {
      "Python parent id": 11765, "Python id": 11766, "finished": true, "Ev Idx": 14666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915671.094, "dur": 0.063,
    "args": {
      "Python parent id": 11765, "Python id": 11767, "finished": true, "Ev Idx": 14667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915671.321, "dur": 0.068,
    "args": {
      "Python parent id": 11765, "Python id": 11768, "finished": true, "Ev Idx": 14668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915671.857, "dur": 0.083,
    "args": {
      "Python parent id": 11765, "Python id": 11769, "finished": true, "Ev Idx": 14669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915672.226, "dur": 0.093,
    "args": {
      "Python parent id": 11765, "Python id": 11770, "finished": true, "Ev Idx": 14670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915672.686, "dur": 8.225,
    "args": {
      "Python parent id": 11765, "Python id": 11771, "finished": true, "Ev Idx": 14671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915672.797, "dur": 0.060,
    "args": {
      "Python parent id": 11771, "Python id": 11772, "finished": true, "Ev Idx": 14672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915673.029, "dur": 0.056,
    "args": {
      "Python parent id": 11771, "Python id": 11773, "finished": true, "Ev Idx": 14673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915673.275, "dur": 0.067,
    "args": {
      "Python parent id": 11771, "Python id": 11774, "finished": true, "Ev Idx": 14674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915673.735, "dur": 0.039,
    "args": {
      "Python parent id": 11771, "Python id": 11775, "finished": true, "Ev Idx": 14675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915674.075, "dur": 6.736,
    "args": {
      "Python parent id": 11771, "Python id": 11776, "finished": true, "Ev Idx": 14676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915674.382, "dur": 6.323,
    "args": {
      "Python parent id": 11776, "Python id": 11777, "finished": true, "Ev Idx": 14677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915674.529, "dur": 3.776,
    "args": {
      "Python parent id": 11777, "Python id": 11778, "finished": true, "Ev Idx": 14678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915674.770, "dur": 3.493,
    "args": {
      "Python parent id": 11778, "Python id": 11779, "finished": true, "Ev Idx": 14679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915674.894, "dur": 0.270,
    "args": {
      "Python parent id": 11779, "Python id": 11780, "finished": true, "Ev Idx": 14680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915675.020, "dur": 0.066,
    "args": {
      "Python parent id": 11780, "Python id": 11781, "finished": true, "Ev Idx": 14681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915675.365, "dur": 2.507,
    "args": {
      "Python parent id": 11779, "Python id": 11782, "finished": true, "Ev Idx": 14682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915675.580, "dur": 2.173,
    "args": {
      "Python parent id": 11782, "Python id": 11783, "finished": true, "Ev Idx": 14683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915675.830, "dur": 1.856,
    "args": {
      "Python parent id": 11783, "Python id": 11784, "finished": true, "Ev Idx": 14684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915676.006, "dur": 1.422,
    "args": {
      "Python parent id": 11784, "Python id": 11785, "finished": true, "Ev Idx": 14685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915676.205, "dur": 0.578,
    "args": {
      "Python parent id": 11785, "Python id": 11786, "finished": true, "Ev Idx": 14686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915676.286, "dur": 0.046,
    "args": {
      "Python parent id": 11786, "Python id": 11787, "finished": true, "Ev Idx": 14687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915676.527, "dur": 0.216,
    "args": {
      "Python parent id": 11786, "Python id": 11788, "finished": true, "Ev Idx": 14688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915678.051, "dur": 0.162,
    "args": {
      "Python parent id": 11779, "Python id": 11789, "finished": true, "Ev Idx": 14689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915678.547, "dur": 0.110,
    "args": {
      "Python parent id": 11777, "Python id": 11790, "finished": true, "Ev Idx": 14690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915678.928, "dur": 1.727,
    "args": {
      "Python parent id": 11777, "Python id": 11791, "finished": true, "Ev Idx": 14691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915679.097, "dur": 1.488,
    "args": {
      "Python parent id": 11791, "Python id": 11792, "finished": true, "Ev Idx": 14692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915679.257, "dur": 0.767,
    "args": {
      "Python parent id": 11792, "Python id": 11793, "finished": true, "Ev Idx": 14693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915679.373, "dur": 0.572,
    "args": {
      "Python parent id": 11793, "Python id": 11794, "finished": true, "Ev Idx": 14694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915679.819, "dur": 0.090,
    "args": {
      "Python parent id": 11794, "Python id": 11795, "finished": true, "Ev Idx": 14695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915680.198, "dur": 0.337,
    "args": {
      "Python parent id": 11792, "Python id": 11796, "finished": true, "Ev Idx": 14696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915681.127, "dur": 0.283,
    "args": {
      "Python parent id": 11761, "Python id": 11797, "finished": true, "Ev Idx": 14697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915682.229, "dur": 1.229,
    "args": {
      "Python parent id": 11761, "Python id": 11798, "finished": true, "Ev Idx": 14698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915682.719, "dur": 0.687,
    "args": {
      "Python parent id": 11798, "Python id": 11799, "finished": true, "Ev Idx": 14699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915684.002, "dur": 3.579,
    "args": {
      "Python parent id": 11760, "Python id": 11800, "finished": true, "Ev Idx": 14700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915688.341, "dur": 7.168,
    "args": {
      "Python parent id": 12, "Python id": 11801, "finished": true, "Ev Idx": 14701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915695.979, "dur": 0.144,
    "args": {
      "Python parent id": 12, "Python id": 11802, "finished": true, "Ev Idx": 14702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915696.820, "dur": 0.177,
    "args": {
      "Python parent id": 12, "Python id": 11803, "finished": true, "Ev Idx": 14703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915697.521, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 11804, "finished": true, "Ev Idx": 14704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915698.013, "dur": 9.989,
    "args": {
      "Python parent id": 12, "Python id": 11805, "finished": true, "Ev Idx": 14705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915709.425, "dur": 16.477,
    "args": {
      "Python parent id": 12, "Python id": 11806, "finished": true, "Ev Idx": 14706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915709.762, "dur": 11.501,
    "args": {
      "Python parent id": 11806, "Python id": 11807, "finished": true, "Ev Idx": 14707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915709.914, "dur": 0.515,
    "args": {
      "Python parent id": 11807, "Python id": 11808, "finished": true, "Ev Idx": 14708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915710.070, "dur": 0.288,
    "args": {
      "Python parent id": 11808, "Python id": 11809, "finished": true, "Ev Idx": 14709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915710.217, "dur": 0.104,
    "args": {
      "Python parent id": 11809, "Python id": 11810, "finished": true, "Ev Idx": 14710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915710.753, "dur": 8.578,
    "args": {
      "Python parent id": 11807, "Python id": 11811, "finished": true, "Ev Idx": 14711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915710.855, "dur": 0.085,
    "args": {
      "Python parent id": 11811, "Python id": 11812, "finished": true, "Ev Idx": 14712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915711.059, "dur": 0.059,
    "args": {
      "Python parent id": 11811, "Python id": 11813, "finished": true, "Ev Idx": 14713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915711.299, "dur": 0.058,
    "args": {
      "Python parent id": 11811, "Python id": 11814, "finished": true, "Ev Idx": 14714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915711.815, "dur": 0.035,
    "args": {
      "Python parent id": 11811, "Python id": 11815, "finished": true, "Ev Idx": 14715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915712.105, "dur": 0.077,
    "args": {
      "Python parent id": 11811, "Python id": 11816, "finished": true, "Ev Idx": 14716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915712.527, "dur": 6.748,
    "args": {
      "Python parent id": 11811, "Python id": 11817, "finished": true, "Ev Idx": 14717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915712.617, "dur": 0.069,
    "args": {
      "Python parent id": 11817, "Python id": 11818, "finished": true, "Ev Idx": 14718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915712.848, "dur": 0.065,
    "args": {
      "Python parent id": 11817, "Python id": 11819, "finished": true, "Ev Idx": 14719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915713.073, "dur": 0.074,
    "args": {
      "Python parent id": 11817, "Python id": 11820, "finished": true, "Ev Idx": 14720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915713.523, "dur": 0.040,
    "args": {
      "Python parent id": 11817, "Python id": 11821, "finished": true, "Ev Idx": 14721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915713.894, "dur": 5.296,
    "args": {
      "Python parent id": 11817, "Python id": 11822, "finished": true, "Ev Idx": 14722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915714.090, "dur": 5.022,
    "args": {
      "Python parent id": 11822, "Python id": 11823, "finished": true, "Ev Idx": 14723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915714.263, "dur": 3.160,
    "args": {
      "Python parent id": 11823, "Python id": 11824, "finished": true, "Ev Idx": 14724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915714.486, "dur": 2.868,
    "args": {
      "Python parent id": 11824, "Python id": 11825, "finished": true, "Ev Idx": 14725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915714.607, "dur": 0.252,
    "args": {
      "Python parent id": 11825, "Python id": 11826, "finished": true, "Ev Idx": 14726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915714.733, "dur": 0.083,
    "args": {
      "Python parent id": 11826, "Python id": 11827, "finished": true, "Ev Idx": 14727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.015, "dur": 2.008,
    "args": {
      "Python parent id": 11825, "Python id": 11828, "finished": true, "Ev Idx": 14728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.221, "dur": 1.714,
    "args": {
      "Python parent id": 11828, "Python id": 11829, "finished": true, "Ev Idx": 14729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.411, "dur": 1.476,
    "args": {
      "Python parent id": 11829, "Python id": 11830, "finished": true, "Ev Idx": 14730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.541, "dur": 1.134,
    "args": {
      "Python parent id": 11830, "Python id": 11831, "finished": true, "Ev Idx": 14731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.693, "dur": 0.502,
    "args": {
      "Python parent id": 11831, "Python id": 11832, "finished": true, "Ev Idx": 14732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915715.774, "dur": 0.050,
    "args": {
      "Python parent id": 11832, "Python id": 11833, "finished": true, "Ev Idx": 14733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915716.023, "dur": 0.120,
    "args": {
      "Python parent id": 11832, "Python id": 11834, "finished": true, "Ev Idx": 14734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915717.208, "dur": 0.091,
    "args": {
      "Python parent id": 11825, "Python id": 11835, "finished": true, "Ev Idx": 14735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915717.613, "dur": 0.079,
    "args": {
      "Python parent id": 11823, "Python id": 11836, "finished": true, "Ev Idx": 14736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915717.940, "dur": 1.129,
    "args": {
      "Python parent id": 11823, "Python id": 11837, "finished": true, "Ev Idx": 14737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915718.075, "dur": 0.949,
    "args": {
      "Python parent id": 11837, "Python id": 11838, "finished": true, "Ev Idx": 14738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915718.242, "dur": 0.363,
    "args": {
      "Python parent id": 11838, "Python id": 11839, "finished": true, "Ev Idx": 14739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915718.365, "dur": 0.180,
    "args": {
      "Python parent id": 11839, "Python id": 11840, "finished": true, "Ev Idx": 14740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915718.459, "dur": 0.049,
    "args": {
      "Python parent id": 11840, "Python id": 11841, "finished": true, "Ev Idx": 14741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915718.739, "dur": 0.251,
    "args": {
      "Python parent id": 11838, "Python id": 11842, "finished": true, "Ev Idx": 14742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915719.462, "dur": 0.139,
    "args": {
      "Python parent id": 11807, "Python id": 11843, "finished": true, "Ev Idx": 14743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915720.254, "dur": 0.873,
    "args": {
      "Python parent id": 11807, "Python id": 11844, "finished": true, "Ev Idx": 14744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915720.657, "dur": 0.414,
    "args": {
      "Python parent id": 11844, "Python id": 11845, "finished": true, "Ev Idx": 14745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915721.527, "dur": 4.327,
    "args": {
      "Python parent id": 11806, "Python id": 11846, "finished": true, "Ev Idx": 14746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915726.864, "dur": 93.183,
    "args": {
      "Python parent id": 12, "Python id": 11847, "finished": true, "Ev Idx": 14747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915727.275, "dur": 0.429,
    "args": {
      "Python parent id": 11847, "Python id": 11848, "finished": true, "Ev Idx": 14748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915727.579, "dur": 0.075,
    "args": {
      "Python parent id": 11848, "Python id": 11849, "finished": true, "Ev Idx": 14749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915728.078, "dur": 0.329,
    "args": {
      "Python parent id": 11847, "Python id": 11850, "finished": true, "Ev Idx": 14750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915728.695, "dur": 0.241,
    "args": {
      "Python parent id": 11847, "Python id": 11851, "finished": true, "Ev Idx": 14751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915729.237, "dur": 1.678,
    "args": {
      "Python parent id": 11847, "Python id": 11852, "finished": true, "Ev Idx": 14752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915731.636, "dur": 88.289,
    "args": {
      "Python parent id": 11847, "Python id": 11853, "finished": true, "Ev Idx": 14753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915732.678, "dur": 87.062,
    "args": {
      "Python parent id": 11853, "Python id": 11854, "finished": true, "Ev Idx": 14754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915733.080, "dur": 86.572,
    "args": {
      "Python parent id": 11854, "Python id": 11855, "finished": true, "Ev Idx": 14755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915738.653, "dur": 79.591,
    "args": {
      "Python parent id": 11855, "Python id": 11856, "finished": true, "Ev Idx": 14756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915738.997, "dur": 79.090,
    "args": {
      "Python parent id": 11856, "Python id": 11857, "finished": true, "Ev Idx": 14757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915739.651, "dur": 78.290,
    "args": {
      "Python parent id": 11857, "Python id": 11858, "finished": true, "Ev Idx": 14758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915740.245, "dur": 0.438,
    "args": {
      "Python parent id": 11858, "Python id": 11859, "finished": true, "Ev Idx": 14759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915740.519, "dur": 0.124,
    "args": {
      "Python parent id": 11859, "Python id": 11860, "finished": true, "Ev Idx": 14760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915741.001, "dur": 0.390,
    "args": {
      "Python parent id": 11858, "Python id": 11861, "finished": true, "Ev Idx": 14761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915741.292, "dur": 0.049,
    "args": {
      "Python parent id": 11861, "Python id": 11862, "finished": true, "Ev Idx": 14762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915741.740, "dur": 76.071,
    "args": {
      "Python parent id": 11858, "Python id": 11863, "finished": true, "Ev Idx": 14763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915742.490, "dur": 74.086,
    "args": {
      "Python parent id": 11863, "Python id": 11864, "finished": true, "Ev Idx": 14764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915749.756, "dur": 18.600,
    "args": {
      "Python parent id": 11864, "Python id": 11865, "finished": true, "Ev Idx": 14765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915749.918, "dur": 0.153,
    "args": {
      "Python parent id": 11865, "Python id": 11866, "finished": true, "Ev Idx": 14766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915750.358, "dur": 17.826,
    "args": {
      "Python parent id": 11865, "Python id": 11867, "finished": true, "Ev Idx": 14767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915750.460, "dur": 7.923,
    "args": {
      "Python parent id": 11867, "Python id": 11868, "finished": true, "Ev Idx": 14768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915775.701, "dur": 7.859,
    "args": {
      "Python parent id": 11864, "Python id": 11869, "finished": true, "Ev Idx": 14769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915792.853, "dur": 8.757,
    "args": {
      "Python parent id": 11864, "Python id": 11870, "finished": true, "Ev Idx": 14770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915821.464, "dur": 17.745,
    "args": {
      "Python parent id": 12, "Python id": 11871, "finished": true, "Ev Idx": 14771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915821.819, "dur": 13.258,
    "args": {
      "Python parent id": 11871, "Python id": 11872, "finished": true, "Ev Idx": 14772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915821.963, "dur": 0.498,
    "args": {
      "Python parent id": 11872, "Python id": 11873, "finished": true, "Ev Idx": 14773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915822.131, "dur": 0.265,
    "args": {
      "Python parent id": 11873, "Python id": 11874, "finished": true, "Ev Idx": 14774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915822.287, "dur": 0.073,
    "args": {
      "Python parent id": 11874, "Python id": 11875, "finished": true, "Ev Idx": 14775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915822.821, "dur": 9.595,
    "args": {
      "Python parent id": 11872, "Python id": 11876, "finished": true, "Ev Idx": 14776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915822.925, "dur": 0.071,
    "args": {
      "Python parent id": 11876, "Python id": 11877, "finished": true, "Ev Idx": 14777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915823.155, "dur": 0.063,
    "args": {
      "Python parent id": 11876, "Python id": 11878, "finished": true, "Ev Idx": 14778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915823.391, "dur": 0.068,
    "args": {
      "Python parent id": 11876, "Python id": 11879, "finished": true, "Ev Idx": 14779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915823.909, "dur": 0.040,
    "args": {
      "Python parent id": 11876, "Python id": 11880, "finished": true, "Ev Idx": 14780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915824.229, "dur": 0.103,
    "args": {
      "Python parent id": 11876, "Python id": 11881, "finished": true, "Ev Idx": 14781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915824.684, "dur": 7.651,
    "args": {
      "Python parent id": 11876, "Python id": 11882, "finished": true, "Ev Idx": 14782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915824.779, "dur": 0.064,
    "args": {
      "Python parent id": 11882, "Python id": 11883, "finished": true, "Ev Idx": 14783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915825.007, "dur": 0.062,
    "args": {
      "Python parent id": 11882, "Python id": 11884, "finished": true, "Ev Idx": 14784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915825.271, "dur": 0.073,
    "args": {
      "Python parent id": 11882, "Python id": 11885, "finished": true, "Ev Idx": 14785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915825.750, "dur": 0.034,
    "args": {
      "Python parent id": 11882, "Python id": 11886, "finished": true, "Ev Idx": 14786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915826.113, "dur": 6.139,
    "args": {
      "Python parent id": 11882, "Python id": 11887, "finished": true, "Ev Idx": 14787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915826.447, "dur": 5.722,
    "args": {
      "Python parent id": 11887, "Python id": 11888, "finished": true, "Ev Idx": 14788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915826.576, "dur": 3.676,
    "args": {
      "Python parent id": 11888, "Python id": 11889, "finished": true, "Ev Idx": 14789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915826.801, "dur": 3.378,
    "args": {
      "Python parent id": 11889, "Python id": 11890, "finished": true, "Ev Idx": 14790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915826.927, "dur": 0.277,
    "args": {
      "Python parent id": 11890, "Python id": 11891, "finished": true, "Ev Idx": 14791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915827.081, "dur": 0.064,
    "args": {
      "Python parent id": 11891, "Python id": 11892, "finished": true, "Ev Idx": 14792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915827.388, "dur": 2.416,
    "args": {
      "Python parent id": 11890, "Python id": 11893, "finished": true, "Ev Idx": 14793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915827.628, "dur": 2.056,
    "args": {
      "Python parent id": 11893, "Python id": 11894, "finished": true, "Ev Idx": 14794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915827.845, "dur": 1.771,
    "args": {
      "Python parent id": 11894, "Python id": 11895, "finished": true, "Ev Idx": 14795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915828.030, "dur": 1.377,
    "args": {
      "Python parent id": 11895, "Python id": 11896, "finished": true, "Ev Idx": 14796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915828.222, "dur": 0.533,
    "args": {
      "Python parent id": 11896, "Python id": 11897, "finished": true, "Ev Idx": 14797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915828.312, "dur": 0.047,
    "args": {
      "Python parent id": 11897, "Python id": 11898, "finished": true, "Ev Idx": 14798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915828.550, "dur": 0.153,
    "args": {
      "Python parent id": 11897, "Python id": 11899, "finished": true, "Ev Idx": 14799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915829.984, "dur": 0.144,
    "args": {
      "Python parent id": 11890, "Python id": 11900, "finished": true, "Ev Idx": 14800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915830.458, "dur": 0.117,
    "args": {
      "Python parent id": 11888, "Python id": 11901, "finished": true, "Ev Idx": 14801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915830.845, "dur": 1.280,
    "args": {
      "Python parent id": 11888, "Python id": 11902, "finished": true, "Ev Idx": 14802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915831.026, "dur": 1.038,
    "args": {
      "Python parent id": 11902, "Python id": 11903, "finished": true, "Ev Idx": 14803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915831.186, "dur": 0.414,
    "args": {
      "Python parent id": 11903, "Python id": 11904, "finished": true, "Ev Idx": 14804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915831.324, "dur": 0.197,
    "args": {
      "Python parent id": 11904, "Python id": 11905, "finished": true, "Ev Idx": 14805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915831.432, "dur": 0.060,
    "args": {
      "Python parent id": 11905, "Python id": 11906, "finished": true, "Ev Idx": 14806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915831.725, "dur": 0.291,
    "args": {
      "Python parent id": 11903, "Python id": 11907, "finished": true, "Ev Idx": 14807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915832.568, "dur": 0.276,
    "args": {
      "Python parent id": 11872, "Python id": 11908, "finished": true, "Ev Idx": 14808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915833.615, "dur": 1.299,
    "args": {
      "Python parent id": 11872, "Python id": 11909, "finished": true, "Ev Idx": 14809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915834.229, "dur": 0.631,
    "args": {
      "Python parent id": 11909, "Python id": 11910, "finished": true, "Ev Idx": 14810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915835.475, "dur": 3.658,
    "args": {
      "Python parent id": 11871, "Python id": 11911, "finished": true, "Ev Idx": 14811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915839.872, "dur": 8.108,
    "args": {
      "Python parent id": 12, "Python id": 11912, "finished": true, "Ev Idx": 14812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740915848.486, "dur": 0.127,
    "args": {
      "Python parent id": 12, "Python id": 11913, "finished": true, "Ev Idx": 14813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915849.241, "dur": 0.196,
    "args": {
      "Python parent id": 12, "Python id": 11914, "finished": true, "Ev Idx": 14814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915849.916, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 11915, "finished": true, "Ev Idx": 14815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740915850.400, "dur": 15.957,
    "args": {
      "Python parent id": 12, "Python id": 11916, "finished": true, "Ev Idx": 14816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740915868.652, "dur": 19.370,
    "args": {
      "Python parent id": 12, "Python id": 11917, "finished": true, "Ev Idx": 14817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740915869.061, "dur": 13.967,
    "args": {
      "Python parent id": 11917, "Python id": 11918, "finished": true, "Ev Idx": 14818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915869.228, "dur": 0.726,
    "args": {
      "Python parent id": 11918, "Python id": 11919, "finished": true, "Ev Idx": 14819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915869.385, "dur": 0.424,
    "args": {
      "Python parent id": 11919, "Python id": 11920, "finished": true, "Ev Idx": 14820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915869.537, "dur": 0.192,
    "args": {
      "Python parent id": 11920, "Python id": 11921, "finished": true, "Ev Idx": 14821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915870.347, "dur": 9.967,
    "args": {
      "Python parent id": 11918, "Python id": 11922, "finished": true, "Ev Idx": 14822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915870.466, "dur": 0.099,
    "args": {
      "Python parent id": 11922, "Python id": 11923, "finished": true, "Ev Idx": 14823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915870.685, "dur": 0.064,
    "args": {
      "Python parent id": 11922, "Python id": 11924, "finished": true, "Ev Idx": 14824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915870.902, "dur": 0.064,
    "args": {
      "Python parent id": 11922, "Python id": 11925, "finished": true, "Ev Idx": 14825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915871.431, "dur": 0.070,
    "args": {
      "Python parent id": 11922, "Python id": 11926, "finished": true, "Ev Idx": 14826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915871.812, "dur": 0.073,
    "args": {
      "Python parent id": 11922, "Python id": 11927, "finished": true, "Ev Idx": 14827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915872.220, "dur": 8.036,
    "args": {
      "Python parent id": 11922, "Python id": 11928, "finished": true, "Ev Idx": 14828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915872.312, "dur": 0.065,
    "args": {
      "Python parent id": 11928, "Python id": 11929, "finished": true, "Ev Idx": 14829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915872.561, "dur": 0.063,
    "args": {
      "Python parent id": 11928, "Python id": 11930, "finished": true, "Ev Idx": 14830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915872.845, "dur": 0.072,
    "args": {
      "Python parent id": 11928, "Python id": 11931, "finished": true, "Ev Idx": 14831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740915873.268, "dur": 0.029,
    "args": {
      "Python parent id": 11928, "Python id": 11932, "finished": true, "Ev Idx": 14832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740915873.623, "dur": 6.542,
    "args": {
      "Python parent id": 11928, "Python id": 11933, "finished": true, "Ev Idx": 14833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740915873.980, "dur": 6.107,
    "args": {
      "Python parent id": 11933, "Python id": 11934, "finished": true, "Ev Idx": 14834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740915874.156, "dur": 3.855,
    "args": {
      "Python parent id": 11934, "Python id": 11935, "finished": true, "Ev Idx": 14835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740915874.364, "dur": 3.569,
    "args": {
      "Python parent id": 11935, "Python id": 11936, "finished": true, "Ev Idx": 14836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740915874.523, "dur": 0.250,
    "args": {
      "Python parent id": 11936, "Python id": 11937, "finished": true, "Ev Idx": 14837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740915874.660, "dur": 0.068,
    "args": {
      "Python parent id": 11937, "Python id": 11938, "finished": true, "Ev Idx": 14838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740915874.957, "dur": 2.563,
    "args": {
      "Python parent id": 11936, "Python id": 11939, "finished": true, "Ev Idx": 14839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740915875.174, "dur": 2.224,
    "args": {
      "Python parent id": 11939, "Python id": 11940, "finished": true, "Ev Idx": 14840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740915875.411, "dur": 1.941,
    "args": {
      "Python parent id": 11940, "Python id": 11941, "finished": true, "Ev Idx": 14841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740915875.566, "dur": 1.455,
    "args": {
      "Python parent id": 11941, "Python id": 11942, "finished": true, "Ev Idx": 14842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740915875.768, "dur": 0.567,
    "args": {
      "Python parent id": 11942, "Python id": 11943, "finished": true, "Ev Idx": 14843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740915875.850, "dur": 0.049,
    "args": {
      "Python parent id": 11943, "Python id": 11944, "finished": true, "Ev Idx": 14844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740915876.095, "dur": 0.182,
    "args": {
      "Python parent id": 11943, "Python id": 11945, "finished": true, "Ev Idx": 14845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740915877.721, "dur": 0.165,
    "args": {
      "Python parent id": 11936, "Python id": 11946, "finished": true, "Ev Idx": 14846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740915878.256, "dur": 0.114,
    "args": {
      "Python parent id": 11934, "Python id": 11947, "finished": true, "Ev Idx": 14847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740915878.704, "dur": 1.338,
    "args": {
      "Python parent id": 11934, "Python id": 11948, "finished": true, "Ev Idx": 14848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740915878.877, "dur": 1.123,
    "args": {
      "Python parent id": 11948, "Python id": 11949, "finished": true, "Ev Idx": 14849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740915879.009, "dur": 0.442,
    "args": {
      "Python parent id": 11949, "Python id": 11950, "finished": true, "Ev Idx": 14850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740915879.164, "dur": 0.214,
    "args": {
      "Python parent id": 11950, "Python id": 11951, "finished": true, "Ev Idx": 14851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740915879.288, "dur": 0.060,
    "args": {
      "Python parent id": 11951, "Python id": 11952, "finished": true, "Ev Idx": 14852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740915879.578, "dur": 0.394,
    "args": {
      "Python parent id": 11949, "Python id": 11953, "finished": true, "Ev Idx": 14853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740915880.437, "dur": 0.288,
    "args": {
      "Python parent id": 11918, "Python id": 11954, "finished": true, "Ev Idx": 14854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740915881.509, "dur": 1.359,
    "args": {
      "Python parent id": 11918, "Python id": 11955, "finished": true, "Ev Idx": 14855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740915882.096, "dur": 0.664,
    "args": {
      "Python parent id": 11955, "Python id": 11956, "finished": true, "Ev Idx": 14856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740915883.409, "dur": 4.524,
    "args": {
      "Python parent id": 11917, "Python id": 11957, "finished": true, "Ev Idx": 14857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740915889.025, "dur": 119.699,
    "args": {
      "Python parent id": 12, "Python id": 11958, "finished": true, "Ev Idx": 14858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915889.450, "dur": 0.526,
    "args": {
      "Python parent id": 11958, "Python id": 11959, "finished": true, "Ev Idx": 14859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915889.829, "dur": 0.081,
    "args": {
      "Python parent id": 11959, "Python id": 11960, "finished": true, "Ev Idx": 14860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915890.312, "dur": 0.466,
    "args": {
      "Python parent id": 11958, "Python id": 11961, "finished": true, "Ev Idx": 14861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740915891.006, "dur": 0.231,
    "args": {
      "Python parent id": 11958, "Python id": 11962, "finished": true, "Ev Idx": 14862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740915891.495, "dur": 3.421,
    "args": {
      "Python parent id": 11958, "Python id": 11963, "finished": true, "Ev Idx": 14863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740915895.556, "dur": 113.098,
    "args": {
      "Python parent id": 11958, "Python id": 11964, "finished": true, "Ev Idx": 14864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740915896.613, "dur": 111.762,
    "args": {
      "Python parent id": 11964, "Python id": 11965, "finished": true, "Ev Idx": 14865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915896.970, "dur": 111.293,
    "args": {
      "Python parent id": 11965, "Python id": 11966, "finished": true, "Ev Idx": 14866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740915905.284, "dur": 100.920,
    "args": {
      "Python parent id": 11966, "Python id": 11967, "finished": true, "Ev Idx": 14867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740915905.654, "dur": 100.490,
    "args": {
      "Python parent id": 11967, "Python id": 11968, "finished": true, "Ev Idx": 14868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740915906.297, "dur": 99.641,
    "args": {
      "Python parent id": 11968, "Python id": 11969, "finished": true, "Ev Idx": 14869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915906.873, "dur": 0.496,
    "args": {
      "Python parent id": 11969, "Python id": 11970, "finished": true, "Ev Idx": 14870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915907.191, "dur": 0.105,
    "args": {
      "Python parent id": 11970, "Python id": 11971, "finished": true, "Ev Idx": 14871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740915907.702, "dur": 0.394,
    "args": {
      "Python parent id": 11969, "Python id": 11972, "finished": true, "Ev Idx": 14872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740915908.000, "dur": 0.048,
    "args": {
      "Python parent id": 11972, "Python id": 11973, "finished": true, "Ev Idx": 14873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915908.442, "dur": 97.388,
    "args": {
      "Python parent id": 11969, "Python id": 11974, "finished": true, "Ev Idx": 14874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740915909.313, "dur": 94.701,
    "args": {
      "Python parent id": 11974, "Python id": 11975, "finished": true, "Ev Idx": 14875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740915918.368, "dur": 25.075,
    "args": {
      "Python parent id": 11975, "Python id": 11976, "finished": true, "Ev Idx": 14876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740915918.529, "dur": 0.130,
    "args": {
      "Python parent id": 11976, "Python id": 11977, "finished": true, "Ev Idx": 14877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740915919.030, "dur": 24.307,
    "args": {
      "Python parent id": 11976, "Python id": 11978, "finished": true, "Ev Idx": 14878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740915919.166, "dur": 11.871,
    "args": {
      "Python parent id": 11978, "Python id": 11979, "finished": true, "Ev Idx": 14879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915952.613, "dur": 8.310,
    "args": {
      "Python parent id": 11975, "Python id": 11980, "finished": true, "Ev Idx": 14880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740915970.559, "dur": 9.501,
    "args": {
      "Python parent id": 11975, "Python id": 11981, "finished": true, "Ev Idx": 14881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916010.607, "dur": 22.155,
    "args": {
      "Python parent id": 12, "Python id": 11982, "finished": true, "Ev Idx": 14882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916011.097, "dur": 15.586,
    "args": {
      "Python parent id": 11982, "Python id": 11983, "finished": true, "Ev Idx": 14883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916011.282, "dur": 0.763,
    "args": {
      "Python parent id": 11983, "Python id": 11984, "finished": true, "Ev Idx": 14884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916011.464, "dur": 0.477,
    "args": {
      "Python parent id": 11984, "Python id": 11985, "finished": true, "Ev Idx": 14885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916011.641, "dur": 0.228,
    "args": {
      "Python parent id": 11985, "Python id": 11986, "finished": true, "Ev Idx": 14886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916012.521, "dur": 10.809,
    "args": {
      "Python parent id": 11983, "Python id": 11987, "finished": true, "Ev Idx": 14887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916012.665, "dur": 0.073,
    "args": {
      "Python parent id": 11987, "Python id": 11988, "finished": true, "Ev Idx": 14888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916012.915, "dur": 0.061,
    "args": {
      "Python parent id": 11987, "Python id": 11989, "finished": true, "Ev Idx": 14889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916013.143, "dur": 0.059,
    "args": {
      "Python parent id": 11987, "Python id": 11990, "finished": true, "Ev Idx": 14890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916013.711, "dur": 0.072,
    "args": {
      "Python parent id": 11987, "Python id": 11991, "finished": true, "Ev Idx": 14891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916014.087, "dur": 0.068,
    "args": {
      "Python parent id": 11987, "Python id": 11992, "finished": true, "Ev Idx": 14892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916014.469, "dur": 8.799,
    "args": {
      "Python parent id": 11987, "Python id": 11993, "finished": true, "Ev Idx": 14893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916014.576, "dur": 0.065,
    "args": {
      "Python parent id": 11993, "Python id": 11994, "finished": true, "Ev Idx": 14894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916014.837, "dur": 0.057,
    "args": {
      "Python parent id": 11993, "Python id": 11995, "finished": true, "Ev Idx": 14895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916015.104, "dur": 0.066,
    "args": {
      "Python parent id": 11993, "Python id": 11996, "finished": true, "Ev Idx": 14896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916015.542, "dur": 0.036,
    "args": {
      "Python parent id": 11993, "Python id": 11997, "finished": true, "Ev Idx": 14897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916015.884, "dur": 7.270,
    "args": {
      "Python parent id": 11993, "Python id": 11998, "finished": true, "Ev Idx": 14898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916016.176, "dur": 6.894,
    "args": {
      "Python parent id": 11998, "Python id": 11999, "finished": true, "Ev Idx": 14899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916016.351, "dur": 4.433,
    "args": {
      "Python parent id": 11999, "Python id": 12000, "finished": true, "Ev Idx": 14900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916016.543, "dur": 4.159,
    "args": {
      "Python parent id": 12000, "Python id": 12001, "finished": true, "Ev Idx": 14901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916016.677, "dur": 0.305,
    "args": {
      "Python parent id": 12001, "Python id": 12002, "finished": true, "Ev Idx": 14902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916016.833, "dur": 0.092,
    "args": {
      "Python parent id": 12002, "Python id": 12003, "finished": true, "Ev Idx": 14903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916017.181, "dur": 3.085,
    "args": {
      "Python parent id": 12001, "Python id": 12004, "finished": true, "Ev Idx": 14904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916017.409, "dur": 2.711,
    "args": {
      "Python parent id": 12004, "Python id": 12005, "finished": true, "Ev Idx": 14905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916017.741, "dur": 2.316,
    "args": {
      "Python parent id": 12005, "Python id": 12006, "finished": true, "Ev Idx": 14906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916017.934, "dur": 1.718,
    "args": {
      "Python parent id": 12006, "Python id": 12007, "finished": true, "Ev Idx": 14907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916018.170, "dur": 0.651,
    "args": {
      "Python parent id": 12007, "Python id": 12008, "finished": true, "Ev Idx": 14908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916018.259, "dur": 0.074,
    "args": {
      "Python parent id": 12008, "Python id": 12009, "finished": true, "Ev Idx": 14909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916018.549, "dur": 0.207,
    "args": {
      "Python parent id": 12008, "Python id": 12010, "finished": true, "Ev Idx": 14910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916020.465, "dur": 0.178,
    "args": {
      "Python parent id": 12001, "Python id": 12011, "finished": true, "Ev Idx": 14911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916021.047, "dur": 0.132,
    "args": {
      "Python parent id": 11999, "Python id": 12012, "finished": true, "Ev Idx": 14912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916021.571, "dur": 1.453,
    "args": {
      "Python parent id": 11999, "Python id": 12013, "finished": true, "Ev Idx": 14913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916021.782, "dur": 1.193,
    "args": {
      "Python parent id": 12013, "Python id": 12014, "finished": true, "Ev Idx": 14914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916021.934, "dur": 0.483,
    "args": {
      "Python parent id": 12014, "Python id": 12015, "finished": true, "Ev Idx": 14915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916022.089, "dur": 0.243,
    "args": {
      "Python parent id": 12015, "Python id": 12016, "finished": true, "Ev Idx": 14916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916022.215, "dur": 0.055,
    "args": {
      "Python parent id": 12016, "Python id": 12017, "finished": true, "Ev Idx": 14917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916022.569, "dur": 0.360,
    "args": {
      "Python parent id": 12014, "Python id": 12018, "finished": true, "Ev Idx": 14918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916023.511, "dur": 0.367,
    "args": {
      "Python parent id": 11983, "Python id": 12019, "finished": true, "Ev Idx": 14919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916024.779, "dur": 1.731,
    "args": {
      "Python parent id": 11983, "Python id": 12020, "finished": true, "Ev Idx": 14920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916025.653, "dur": 0.768,
    "args": {
      "Python parent id": 12020, "Python id": 12021, "finished": true, "Ev Idx": 14921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916027.151, "dur": 5.502,
    "args": {
      "Python parent id": 11982, "Python id": 12022, "finished": true, "Ev Idx": 14922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916033.448, "dur": 9.473,
    "args": {
      "Python parent id": 12, "Python id": 12023, "finished": true, "Ev Idx": 14923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916043.421, "dur": 0.112,
    "args": {
      "Python parent id": 12, "Python id": 12024, "finished": true, "Ev Idx": 14924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916044.332, "dur": 0.218,
    "args": {
      "Python parent id": 12, "Python id": 12025, "finished": true, "Ev Idx": 14925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916045.080, "dur": 0.074,
    "args": {
      "Python parent id": 12, "Python id": 12026, "finished": true, "Ev Idx": 14926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916045.591, "dur": 12.565,
    "args": {
      "Python parent id": 12, "Python id": 12027, "finished": true, "Ev Idx": 14927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916059.602, "dur": 17.032,
    "args": {
      "Python parent id": 12, "Python id": 12028, "finished": true, "Ev Idx": 14928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916060.006, "dur": 12.860,
    "args": {
      "Python parent id": 12028, "Python id": 12029, "finished": true, "Ev Idx": 14929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916060.171, "dur": 0.628,
    "args": {
      "Python parent id": 12029, "Python id": 12030, "finished": true, "Ev Idx": 14930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916060.364, "dur": 0.318,
    "args": {
      "Python parent id": 12030, "Python id": 12031, "finished": true, "Ev Idx": 14931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916060.522, "dur": 0.124,
    "args": {
      "Python parent id": 12031, "Python id": 12032, "finished": true, "Ev Idx": 14932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916061.160, "dur": 9.398,
    "args": {
      "Python parent id": 12029, "Python id": 12033, "finished": true, "Ev Idx": 14933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916061.278, "dur": 0.076,
    "args": {
      "Python parent id": 12033, "Python id": 12034, "finished": true, "Ev Idx": 14934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916061.512, "dur": 0.063,
    "args": {
      "Python parent id": 12033, "Python id": 12035, "finished": true, "Ev Idx": 14935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916061.751, "dur": 0.064,
    "args": {
      "Python parent id": 12033, "Python id": 12036, "finished": true, "Ev Idx": 14936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916062.234, "dur": 0.043,
    "args": {
      "Python parent id": 12033, "Python id": 12037, "finished": true, "Ev Idx": 14937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916062.607, "dur": 0.076,
    "args": {
      "Python parent id": 12033, "Python id": 12038, "finished": true, "Ev Idx": 14938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916063.016, "dur": 7.474,
    "args": {
      "Python parent id": 12033, "Python id": 12039, "finished": true, "Ev Idx": 14939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916063.108, "dur": 0.061,
    "args": {
      "Python parent id": 12039, "Python id": 12040, "finished": true, "Ev Idx": 14940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916063.345, "dur": 0.063,
    "args": {
      "Python parent id": 12039, "Python id": 12041, "finished": true, "Ev Idx": 14941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916063.592, "dur": 0.071,
    "args": {
      "Python parent id": 12039, "Python id": 12042, "finished": true, "Ev Idx": 14942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916064.010, "dur": 0.047,
    "args": {
      "Python parent id": 12039, "Python id": 12043, "finished": true, "Ev Idx": 14943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916064.366, "dur": 6.031,
    "args": {
      "Python parent id": 12039, "Python id": 12044, "finished": true, "Ev Idx": 14944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916064.581, "dur": 5.733,
    "args": {
      "Python parent id": 12044, "Python id": 12045, "finished": true, "Ev Idx": 14945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916064.718, "dur": 3.608,
    "args": {
      "Python parent id": 12045, "Python id": 12046, "finished": true, "Ev Idx": 14946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916064.917, "dur": 3.333,
    "args": {
      "Python parent id": 12046, "Python id": 12047, "finished": true, "Ev Idx": 14947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916065.054, "dur": 0.293,
    "args": {
      "Python parent id": 12047, "Python id": 12048, "finished": true, "Ev Idx": 14948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916065.206, "dur": 0.088,
    "args": {
      "Python parent id": 12048, "Python id": 12049, "finished": true, "Ev Idx": 14949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916065.526, "dur": 2.421,
    "args": {
      "Python parent id": 12047, "Python id": 12050, "finished": true, "Ev Idx": 14950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916065.698, "dur": 2.153,
    "args": {
      "Python parent id": 12050, "Python id": 12051, "finished": true, "Ev Idx": 14951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916065.898, "dur": 1.881,
    "args": {
      "Python parent id": 12051, "Python id": 12052, "finished": true, "Ev Idx": 14952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916066.063, "dur": 1.357,
    "args": {
      "Python parent id": 12052, "Python id": 12053, "finished": true, "Ev Idx": 14953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916066.267, "dur": 0.541,
    "args": {
      "Python parent id": 12053, "Python id": 12054, "finished": true, "Ev Idx": 14954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916066.357, "dur": 0.043,
    "args": {
      "Python parent id": 12054, "Python id": 12055, "finished": true, "Ev Idx": 14955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916066.572, "dur": 0.188,
    "args": {
      "Python parent id": 12054, "Python id": 12056, "finished": true, "Ev Idx": 14956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916068.132, "dur": 0.083,
    "args": {
      "Python parent id": 12047, "Python id": 12057, "finished": true, "Ev Idx": 14957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916068.528, "dur": 0.128,
    "args": {
      "Python parent id": 12045, "Python id": 12058, "finished": true, "Ev Idx": 14958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916068.951, "dur": 1.317,
    "args": {
      "Python parent id": 12045, "Python id": 12059, "finished": true, "Ev Idx": 14959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916069.135, "dur": 1.088,
    "args": {
      "Python parent id": 12059, "Python id": 12060, "finished": true, "Ev Idx": 14960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916069.305, "dur": 0.415,
    "args": {
      "Python parent id": 12060, "Python id": 12061, "finished": true, "Ev Idx": 14961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916069.447, "dur": 0.214,
    "args": {
      "Python parent id": 12061, "Python id": 12062, "finished": true, "Ev Idx": 14962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916069.553, "dur": 0.074,
    "args": {
      "Python parent id": 12062, "Python id": 12063, "finished": true, "Ev Idx": 14963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916069.856, "dur": 0.323,
    "args": {
      "Python parent id": 12060, "Python id": 12064, "finished": true, "Ev Idx": 14964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916070.694, "dur": 0.198,
    "args": {
      "Python parent id": 12029, "Python id": 12065, "finished": true, "Ev Idx": 14965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916071.615, "dur": 1.115,
    "args": {
      "Python parent id": 12029, "Python id": 12066, "finished": true, "Ev Idx": 14966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916072.134, "dur": 0.546,
    "args": {
      "Python parent id": 12066, "Python id": 12067, "finished": true, "Ev Idx": 14967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916073.242, "dur": 3.316,
    "args": {
      "Python parent id": 12028, "Python id": 12068, "finished": true, "Ev Idx": 14968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916077.666, "dur": 100.338,
    "args": {
      "Python parent id": 12, "Python id": 12069, "finished": true, "Ev Idx": 14969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916078.134, "dur": 0.532,
    "args": {
      "Python parent id": 12069, "Python id": 12070, "finished": true, "Ev Idx": 14970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916078.506, "dur": 0.094,
    "args": {
      "Python parent id": 12070, "Python id": 12071, "finished": true, "Ev Idx": 14971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916079.010, "dur": 0.570,
    "args": {
      "Python parent id": 12069, "Python id": 12072, "finished": true, "Ev Idx": 14972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916079.818, "dur": 0.252,
    "args": {
      "Python parent id": 12069, "Python id": 12073, "finished": true, "Ev Idx": 14973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916080.358, "dur": 1.976,
    "args": {
      "Python parent id": 12069, "Python id": 12074, "finished": true, "Ev Idx": 14974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916083.025, "dur": 94.906,
    "args": {
      "Python parent id": 12069, "Python id": 12075, "finished": true, "Ev Idx": 14975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916084.097, "dur": 93.619,
    "args": {
      "Python parent id": 12075, "Python id": 12076, "finished": true, "Ev Idx": 14976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916084.456, "dur": 93.201,
    "args": {
      "Python parent id": 12076, "Python id": 12077, "finished": true, "Ev Idx": 14977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916090.936, "dur": 85.194,
    "args": {
      "Python parent id": 12077, "Python id": 12078, "finished": true, "Ev Idx": 14978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916091.311, "dur": 84.736,
    "args": {
      "Python parent id": 12078, "Python id": 12079, "finished": true, "Ev Idx": 14979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916091.922, "dur": 83.942,
    "args": {
      "Python parent id": 12079, "Python id": 12080, "finished": true, "Ev Idx": 14980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916092.500, "dur": 0.512,
    "args": {
      "Python parent id": 12080, "Python id": 12081, "finished": true, "Ev Idx": 14981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916092.839, "dur": 0.103,
    "args": {
      "Python parent id": 12081, "Python id": 12082, "finished": true, "Ev Idx": 14982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916093.347, "dur": 0.329,
    "args": {
      "Python parent id": 12080, "Python id": 12083, "finished": true, "Ev Idx": 14983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916093.572, "dur": 0.053,
    "args": {
      "Python parent id": 12083, "Python id": 12084, "finished": true, "Ev Idx": 14984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916094.002, "dur": 81.694,
    "args": {
      "Python parent id": 12080, "Python id": 12085, "finished": true, "Ev Idx": 14985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916094.785, "dur": 79.545,
    "args": {
      "Python parent id": 12085, "Python id": 12086, "finished": true, "Ev Idx": 14986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916103.042, "dur": 20.674,
    "args": {
      "Python parent id": 12086, "Python id": 12087, "finished": true, "Ev Idx": 14987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916103.205, "dur": 0.144,
    "args": {
      "Python parent id": 12087, "Python id": 12088, "finished": true, "Ev Idx": 14988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916103.711, "dur": 19.814,
    "args": {
      "Python parent id": 12087, "Python id": 12089, "finished": true, "Ev Idx": 14989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916103.880, "dur": 8.263,
    "args": {
      "Python parent id": 12089, "Python id": 12090, "finished": true, "Ev Idx": 14990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916132.191, "dur": 8.018,
    "args": {
      "Python parent id": 12086, "Python id": 12091, "finished": true, "Ev Idx": 14991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916148.991, "dur": 8.927,
    "args": {
      "Python parent id": 12086, "Python id": 12092, "finished": true, "Ev Idx": 14992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916179.361, "dur": 18.991,
    "args": {
      "Python parent id": 12, "Python id": 12093, "finished": true, "Ev Idx": 14993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916179.764, "dur": 14.445,
    "args": {
      "Python parent id": 12093, "Python id": 12094, "finished": true, "Ev Idx": 14994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916179.967, "dur": 0.656,
    "args": {
      "Python parent id": 12094, "Python id": 12095, "finished": true, "Ev Idx": 14995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916180.199, "dur": 0.313,
    "args": {
      "Python parent id": 12095, "Python id": 12096, "finished": true, "Ev Idx": 14996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916180.359, "dur": 0.121,
    "args": {
      "Python parent id": 12096, "Python id": 12097, "finished": true, "Ev Idx": 14997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916181.099, "dur": 10.274,
    "args": {
      "Python parent id": 12094, "Python id": 12098, "finished": true, "Ev Idx": 14998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916181.207, "dur": 0.118,
    "args": {
      "Python parent id": 12098, "Python id": 12099, "finished": true, "Ev Idx": 14999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916181.491, "dur": 0.062,
    "args": {
      "Python parent id": 12098, "Python id": 12100, "finished": true, "Ev Idx": 15000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916181.726, "dur": 0.059,
    "args": {
      "Python parent id": 12098, "Python id": 12101, "finished": true, "Ev Idx": 15001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916182.241, "dur": 0.050,
    "args": {
      "Python parent id": 12098, "Python id": 12102, "finished": true, "Ev Idx": 15002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916182.635, "dur": 0.092,
    "args": {
      "Python parent id": 12098, "Python id": 12103, "finished": true, "Ev Idx": 15003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916183.064, "dur": 8.227,
    "args": {
      "Python parent id": 12098, "Python id": 12104, "finished": true, "Ev Idx": 15004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916183.200, "dur": 0.064,
    "args": {
      "Python parent id": 12104, "Python id": 12105, "finished": true, "Ev Idx": 15005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916183.441, "dur": 0.062,
    "args": {
      "Python parent id": 12104, "Python id": 12106, "finished": true, "Ev Idx": 15006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916183.711, "dur": 0.071,
    "args": {
      "Python parent id": 12104, "Python id": 12107, "finished": true, "Ev Idx": 15007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916184.177, "dur": 0.029,
    "args": {
      "Python parent id": 12104, "Python id": 12108, "finished": true, "Ev Idx": 15008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916184.527, "dur": 6.619,
    "args": {
      "Python parent id": 12104, "Python id": 12109, "finished": true, "Ev Idx": 15009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916184.888, "dur": 6.179,
    "args": {
      "Python parent id": 12109, "Python id": 12110, "finished": true, "Ev Idx": 15010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916185.020, "dur": 3.911,
    "args": {
      "Python parent id": 12110, "Python id": 12111, "finished": true, "Ev Idx": 15011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916185.218, "dur": 3.636,
    "args": {
      "Python parent id": 12111, "Python id": 12112, "finished": true, "Ev Idx": 15012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916185.344, "dur": 0.279,
    "args": {
      "Python parent id": 12112, "Python id": 12113, "finished": true, "Ev Idx": 15013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916185.485, "dur": 0.094,
    "args": {
      "Python parent id": 12113, "Python id": 12114, "finished": true, "Ev Idx": 15014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916185.861, "dur": 2.604,
    "args": {
      "Python parent id": 12112, "Python id": 12115, "finished": true, "Ev Idx": 15015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916186.111, "dur": 2.237,
    "args": {
      "Python parent id": 12115, "Python id": 12116, "finished": true, "Ev Idx": 15016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916186.363, "dur": 1.913,
    "args": {
      "Python parent id": 12116, "Python id": 12117, "finished": true, "Ev Idx": 15017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916186.531, "dur": 1.460,
    "args": {
      "Python parent id": 12117, "Python id": 12118, "finished": true, "Ev Idx": 15018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916186.733, "dur": 0.680,
    "args": {
      "Python parent id": 12118, "Python id": 12119, "finished": true, "Ev Idx": 15019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916186.824, "dur": 0.053,
    "args": {
      "Python parent id": 12119, "Python id": 12120, "finished": true, "Ev Idx": 15020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916187.090, "dur": 0.237,
    "args": {
      "Python parent id": 12119, "Python id": 12121, "finished": true, "Ev Idx": 15021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916188.637, "dur": 0.157,
    "args": {
      "Python parent id": 12112, "Python id": 12122, "finished": true, "Ev Idx": 15022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916189.197, "dur": 0.119,
    "args": {
      "Python parent id": 12110, "Python id": 12123, "finished": true, "Ev Idx": 15023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916189.669, "dur": 1.357,
    "args": {
      "Python parent id": 12110, "Python id": 12124, "finished": true, "Ev Idx": 15024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916189.864, "dur": 1.094,
    "args": {
      "Python parent id": 12124, "Python id": 12125, "finished": true, "Ev Idx": 15025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916190.012, "dur": 0.472,
    "args": {
      "Python parent id": 12125, "Python id": 12126, "finished": true, "Ev Idx": 15026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916190.185, "dur": 0.208,
    "args": {
      "Python parent id": 12126, "Python id": 12127, "finished": true, "Ev Idx": 15027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916190.306, "dur": 0.053,
    "args": {
      "Python parent id": 12127, "Python id": 12128, "finished": true, "Ev Idx": 15028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916190.611, "dur": 0.302,
    "args": {
      "Python parent id": 12125, "Python id": 12129, "finished": true, "Ev Idx": 15029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916191.534, "dur": 0.312,
    "args": {
      "Python parent id": 12094, "Python id": 12130, "finished": true, "Ev Idx": 15030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916192.641, "dur": 1.418,
    "args": {
      "Python parent id": 12094, "Python id": 12131, "finished": true, "Ev Idx": 15031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916193.317, "dur": 0.635,
    "args": {
      "Python parent id": 12131, "Python id": 12132, "finished": true, "Ev Idx": 15032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916194.598, "dur": 3.659,
    "args": {
      "Python parent id": 12093, "Python id": 12133, "finished": true, "Ev Idx": 15033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916198.990, "dur": 8.050,
    "args": {
      "Python parent id": 12, "Python id": 12134, "finished": true, "Ev Idx": 15034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916207.590, "dur": 0.092,
    "args": {
      "Python parent id": 12, "Python id": 12135, "finished": true, "Ev Idx": 15035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916208.346, "dur": 0.216,
    "args": {
      "Python parent id": 12, "Python id": 12136, "finished": true, "Ev Idx": 15036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916209.035, "dur": 0.085,
    "args": {
      "Python parent id": 12, "Python id": 12137, "finished": true, "Ev Idx": 15037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916209.524, "dur": 11.669,
    "args": {
      "Python parent id": 12, "Python id": 12138, "finished": true, "Ev Idx": 15038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916222.494, "dur": 16.323,
    "args": {
      "Python parent id": 12, "Python id": 12139, "finished": true, "Ev Idx": 15039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916222.825, "dur": 12.381,
    "args": {
      "Python parent id": 12139, "Python id": 12140, "finished": true, "Ev Idx": 15040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916223.022, "dur": 0.560,
    "args": {
      "Python parent id": 12140, "Python id": 12141, "finished": true, "Ev Idx": 15041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916223.182, "dur": 0.334,
    "args": {
      "Python parent id": 12141, "Python id": 12142, "finished": true, "Ev Idx": 15042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916223.333, "dur": 0.123,
    "args": {
      "Python parent id": 12142, "Python id": 12143, "finished": true, "Ev Idx": 15043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916223.886, "dur": 9.136,
    "args": {
      "Python parent id": 12140, "Python id": 12144, "finished": true, "Ev Idx": 15044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916223.987, "dur": 0.066,
    "args": {
      "Python parent id": 12144, "Python id": 12145, "finished": true, "Ev Idx": 15045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916224.208, "dur": 0.062,
    "args": {
      "Python parent id": 12144, "Python id": 12146, "finished": true, "Ev Idx": 15046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916224.421, "dur": 0.059,
    "args": {
      "Python parent id": 12144, "Python id": 12147, "finished": true, "Ev Idx": 15047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916224.872, "dur": 0.048,
    "args": {
      "Python parent id": 12144, "Python id": 12148, "finished": true, "Ev Idx": 15048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916225.227, "dur": 0.080,
    "args": {
      "Python parent id": 12144, "Python id": 12149, "finished": true, "Ev Idx": 15049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916225.660, "dur": 7.304,
    "args": {
      "Python parent id": 12144, "Python id": 12150, "finished": true, "Ev Idx": 15050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916225.786, "dur": 0.064,
    "args": {
      "Python parent id": 12150, "Python id": 12151, "finished": true, "Ev Idx": 15051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916225.993, "dur": 0.064,
    "args": {
      "Python parent id": 12150, "Python id": 12152, "finished": true, "Ev Idx": 15052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916226.274, "dur": 0.078,
    "args": {
      "Python parent id": 12150, "Python id": 12153, "finished": true, "Ev Idx": 15053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916226.746, "dur": 0.052,
    "args": {
      "Python parent id": 12150, "Python id": 12154, "finished": true, "Ev Idx": 15054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916227.140, "dur": 5.746,
    "args": {
      "Python parent id": 12150, "Python id": 12155, "finished": true, "Ev Idx": 15055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916227.424, "dur": 5.387,
    "args": {
      "Python parent id": 12155, "Python id": 12156, "finished": true, "Ev Idx": 15056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916227.579, "dur": 3.296,
    "args": {
      "Python parent id": 12156, "Python id": 12157, "finished": true, "Ev Idx": 15057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916227.790, "dur": 3.018,
    "args": {
      "Python parent id": 12157, "Python id": 12158, "finished": true, "Ev Idx": 15058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916227.974, "dur": 0.295,
    "args": {
      "Python parent id": 12158, "Python id": 12159, "finished": true, "Ev Idx": 15059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916228.140, "dur": 0.086,
    "args": {
      "Python parent id": 12159, "Python id": 12160, "finished": true, "Ev Idx": 15060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916228.437, "dur": 2.077,
    "args": {
      "Python parent id": 12158, "Python id": 12161, "finished": true, "Ev Idx": 15061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916228.608, "dur": 1.819,
    "args": {
      "Python parent id": 12161, "Python id": 12162, "finished": true, "Ev Idx": 15062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916228.817, "dur": 1.561,
    "args": {
      "Python parent id": 12162, "Python id": 12163, "finished": true, "Ev Idx": 15063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916228.966, "dur": 1.168,
    "args": {
      "Python parent id": 12163, "Python id": 12164, "finished": true, "Ev Idx": 15064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916229.167, "dur": 0.519,
    "args": {
      "Python parent id": 12164, "Python id": 12165, "finished": true, "Ev Idx": 15065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916229.271, "dur": 0.051,
    "args": {
      "Python parent id": 12165, "Python id": 12166, "finished": true, "Ev Idx": 15066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916229.503, "dur": 0.130,
    "args": {
      "Python parent id": 12165, "Python id": 12167, "finished": true, "Ev Idx": 15067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916230.701, "dur": 0.071,
    "args": {
      "Python parent id": 12158, "Python id": 12168, "finished": true, "Ev Idx": 15068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916231.092, "dur": 0.133,
    "args": {
      "Python parent id": 12156, "Python id": 12169, "finished": true, "Ev Idx": 15069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916231.542, "dur": 1.216,
    "args": {
      "Python parent id": 12156, "Python id": 12170, "finished": true, "Ev Idx": 15070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916231.716, "dur": 0.997,
    "args": {
      "Python parent id": 12170, "Python id": 12171, "finished": true, "Ev Idx": 15071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916231.851, "dur": 0.409,
    "args": {
      "Python parent id": 12171, "Python id": 12172, "finished": true, "Ev Idx": 15072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916231.988, "dur": 0.214,
    "args": {
      "Python parent id": 12172, "Python id": 12173, "finished": true, "Ev Idx": 15073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916232.102, "dur": 0.068,
    "args": {
      "Python parent id": 12173, "Python id": 12174, "finished": true, "Ev Idx": 15074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916232.390, "dur": 0.279,
    "args": {
      "Python parent id": 12171, "Python id": 12175, "finished": true, "Ev Idx": 15075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916233.175, "dur": 0.195,
    "args": {
      "Python parent id": 12140, "Python id": 12176, "finished": true, "Ev Idx": 15076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916234.027, "dur": 1.055,
    "args": {
      "Python parent id": 12140, "Python id": 12177, "finished": true, "Ev Idx": 15077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916234.490, "dur": 0.540,
    "args": {
      "Python parent id": 12177, "Python id": 12178, "finished": true, "Ev Idx": 15078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916235.474, "dur": 3.282,
    "args": {
      "Python parent id": 12139, "Python id": 12179, "finished": true, "Ev Idx": 15079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916239.770, "dur": 93.520,
    "args": {
      "Python parent id": 12, "Python id": 12180, "finished": true, "Ev Idx": 15080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916240.164, "dur": 0.457,
    "args": {
      "Python parent id": 12180, "Python id": 12181, "finished": true, "Ev Idx": 15081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916240.469, "dur": 0.088,
    "args": {
      "Python parent id": 12181, "Python id": 12182, "finished": true, "Ev Idx": 15082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916240.962, "dur": 0.527,
    "args": {
      "Python parent id": 12180, "Python id": 12183, "finished": true, "Ev Idx": 15083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916241.709, "dur": 0.265,
    "args": {
      "Python parent id": 12180, "Python id": 12184, "finished": true, "Ev Idx": 15084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916242.220, "dur": 1.792,
    "args": {
      "Python parent id": 12180, "Python id": 12185, "finished": true, "Ev Idx": 15085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916244.711, "dur": 88.446,
    "args": {
      "Python parent id": 12180, "Python id": 12186, "finished": true, "Ev Idx": 15086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916245.721, "dur": 87.195,
    "args": {
      "Python parent id": 12186, "Python id": 12187, "finished": true, "Ev Idx": 15087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916246.110, "dur": 86.744,
    "args": {
      "Python parent id": 12187, "Python id": 12188, "finished": true, "Ev Idx": 15088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916251.764, "dur": 79.671,
    "args": {
      "Python parent id": 12188, "Python id": 12189, "finished": true, "Ev Idx": 15089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916252.144, "dur": 79.203,
    "args": {
      "Python parent id": 12189, "Python id": 12190, "finished": true, "Ev Idx": 15090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916252.726, "dur": 78.485,
    "args": {
      "Python parent id": 12190, "Python id": 12191, "finished": true, "Ev Idx": 15091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916253.275, "dur": 0.433,
    "args": {
      "Python parent id": 12191, "Python id": 12192, "finished": true, "Ev Idx": 15092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916253.555, "dur": 0.105,
    "args": {
      "Python parent id": 12192, "Python id": 12193, "finished": true, "Ev Idx": 15093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916254.067, "dur": 0.389,
    "args": {
      "Python parent id": 12191, "Python id": 12194, "finished": true, "Ev Idx": 15094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916254.353, "dur": 0.055,
    "args": {
      "Python parent id": 12194, "Python id": 12195, "finished": true, "Ev Idx": 15095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916254.826, "dur": 76.243,
    "args": {
      "Python parent id": 12191, "Python id": 12196, "finished": true, "Ev Idx": 15096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916255.536, "dur": 74.160,
    "args": {
      "Python parent id": 12196, "Python id": 12197, "finished": true, "Ev Idx": 15097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916263.102, "dur": 19.408,
    "args": {
      "Python parent id": 12197, "Python id": 12198, "finished": true, "Ev Idx": 15098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916263.252, "dur": 0.132,
    "args": {
      "Python parent id": 12198, "Python id": 12199, "finished": true, "Ev Idx": 15099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916263.725, "dur": 18.656,
    "args": {
      "Python parent id": 12198, "Python id": 12200, "finished": true, "Ev Idx": 15100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916263.857, "dur": 8.646,
    "args": {
      "Python parent id": 12200, "Python id": 12201, "finished": true, "Ev Idx": 15101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916290.179, "dur": 7.283,
    "args": {
      "Python parent id": 12197, "Python id": 12202, "finished": true, "Ev Idx": 15102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916306.553, "dur": 8.365,
    "args": {
      "Python parent id": 12197, "Python id": 12203, "finished": true, "Ev Idx": 15103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916334.685, "dur": 18.214,
    "args": {
      "Python parent id": 12, "Python id": 12204, "finished": true, "Ev Idx": 15104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916335.098, "dur": 13.646,
    "args": {
      "Python parent id": 12204, "Python id": 12205, "finished": true, "Ev Idx": 15105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916335.277, "dur": 0.595,
    "args": {
      "Python parent id": 12205, "Python id": 12206, "finished": true, "Ev Idx": 15106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916335.426, "dur": 0.338,
    "args": {
      "Python parent id": 12206, "Python id": 12207, "finished": true, "Ev Idx": 15107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916335.587, "dur": 0.139,
    "args": {
      "Python parent id": 12207, "Python id": 12208, "finished": true, "Ev Idx": 15108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916336.253, "dur": 9.783,
    "args": {
      "Python parent id": 12205, "Python id": 12209, "finished": true, "Ev Idx": 15109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916336.393, "dur": 0.115,
    "args": {
      "Python parent id": 12209, "Python id": 12210, "finished": true, "Ev Idx": 15110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916336.631, "dur": 0.064,
    "args": {
      "Python parent id": 12209, "Python id": 12211, "finished": true, "Ev Idx": 15111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916336.861, "dur": 0.065,
    "args": {
      "Python parent id": 12209, "Python id": 12212, "finished": true, "Ev Idx": 15112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916337.401, "dur": 0.046,
    "args": {
      "Python parent id": 12209, "Python id": 12213, "finished": true, "Ev Idx": 15113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916337.750, "dur": 0.100,
    "args": {
      "Python parent id": 12209, "Python id": 12214, "finished": true, "Ev Idx": 15114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916338.255, "dur": 7.691,
    "args": {
      "Python parent id": 12209, "Python id": 12215, "finished": true, "Ev Idx": 15115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916338.347, "dur": 0.063,
    "args": {
      "Python parent id": 12215, "Python id": 12216, "finished": true, "Ev Idx": 15116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916338.565, "dur": 0.057,
    "args": {
      "Python parent id": 12215, "Python id": 12217, "finished": true, "Ev Idx": 15117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916338.838, "dur": 0.070,
    "args": {
      "Python parent id": 12215, "Python id": 12218, "finished": true, "Ev Idx": 15118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916339.257, "dur": 0.030,
    "args": {
      "Python parent id": 12215, "Python id": 12219, "finished": true, "Ev Idx": 15119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916339.574, "dur": 6.273,
    "args": {
      "Python parent id": 12215, "Python id": 12220, "finished": true, "Ev Idx": 15120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916339.885, "dur": 5.898,
    "args": {
      "Python parent id": 12220, "Python id": 12221, "finished": true, "Ev Idx": 15121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916340.027, "dur": 3.772,
    "args": {
      "Python parent id": 12221, "Python id": 12222, "finished": true, "Ev Idx": 15122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916340.279, "dur": 3.438,
    "args": {
      "Python parent id": 12222, "Python id": 12223, "finished": true, "Ev Idx": 15123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916340.405, "dur": 0.305,
    "args": {
      "Python parent id": 12223, "Python id": 12224, "finished": true, "Ev Idx": 15124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916340.567, "dur": 0.066,
    "args": {
      "Python parent id": 12224, "Python id": 12225, "finished": true, "Ev Idx": 15125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916340.889, "dur": 2.442,
    "args": {
      "Python parent id": 12223, "Python id": 12226, "finished": true, "Ev Idx": 15126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916341.142, "dur": 2.076,
    "args": {
      "Python parent id": 12226, "Python id": 12227, "finished": true, "Ev Idx": 15127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916341.340, "dur": 1.814,
    "args": {
      "Python parent id": 12227, "Python id": 12228, "finished": true, "Ev Idx": 15128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916341.531, "dur": 1.399,
    "args": {
      "Python parent id": 12228, "Python id": 12229, "finished": true, "Ev Idx": 15129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916341.796, "dur": 0.602,
    "args": {
      "Python parent id": 12229, "Python id": 12230, "finished": true, "Ev Idx": 15130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916341.890, "dur": 0.057,
    "args": {
      "Python parent id": 12230, "Python id": 12231, "finished": true, "Ev Idx": 15131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916342.143, "dur": 0.184,
    "args": {
      "Python parent id": 12230, "Python id": 12232, "finished": true, "Ev Idx": 15132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916343.535, "dur": 0.151,
    "args": {
      "Python parent id": 12223, "Python id": 12233, "finished": true, "Ev Idx": 15133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916344.005, "dur": 0.131,
    "args": {
      "Python parent id": 12221, "Python id": 12234, "finished": true, "Ev Idx": 15134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916344.470, "dur": 1.265,
    "args": {
      "Python parent id": 12221, "Python id": 12235, "finished": true, "Ev Idx": 15135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916344.656, "dur": 1.014,
    "args": {
      "Python parent id": 12235, "Python id": 12236, "finished": true, "Ev Idx": 15136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916344.798, "dur": 0.404,
    "args": {
      "Python parent id": 12236, "Python id": 12237, "finished": true, "Ev Idx": 15137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916344.935, "dur": 0.202,
    "args": {
      "Python parent id": 12237, "Python id": 12238, "finished": true, "Ev Idx": 15138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916345.049, "dur": 0.053,
    "args": {
      "Python parent id": 12238, "Python id": 12239, "finished": true, "Ev Idx": 15139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916345.343, "dur": 0.276,
    "args": {
      "Python parent id": 12236, "Python id": 12240, "finished": true, "Ev Idx": 15140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916346.191, "dur": 0.275,
    "args": {
      "Python parent id": 12205, "Python id": 12241, "finished": true, "Ev Idx": 15141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916347.276, "dur": 1.315,
    "args": {
      "Python parent id": 12205, "Python id": 12242, "finished": true, "Ev Idx": 15142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916347.889, "dur": 0.631,
    "args": {
      "Python parent id": 12242, "Python id": 12243, "finished": true, "Ev Idx": 15143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916349.066, "dur": 3.746,
    "args": {
      "Python parent id": 12204, "Python id": 12244, "finished": true, "Ev Idx": 15144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916353.533, "dur": 7.899,
    "args": {
      "Python parent id": 12, "Python id": 12245, "finished": true, "Ev Idx": 15145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916362.006, "dur": 0.088,
    "args": {
      "Python parent id": 12, "Python id": 12246, "finished": true, "Ev Idx": 15146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916362.735, "dur": 0.201,
    "args": {
      "Python parent id": 12, "Python id": 12247, "finished": true, "Ev Idx": 15147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916363.443, "dur": 0.076,
    "args": {
      "Python parent id": 12, "Python id": 12248, "finished": true, "Ev Idx": 15148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916363.920, "dur": 10.199,
    "args": {
      "Python parent id": 12, "Python id": 12249, "finished": true, "Ev Idx": 15149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916375.732, "dur": 16.641,
    "args": {
      "Python parent id": 12, "Python id": 12250, "finished": true, "Ev Idx": 15150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916376.120, "dur": 12.088,
    "args": {
      "Python parent id": 12250, "Python id": 12251, "finished": true, "Ev Idx": 15151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916376.332, "dur": 0.556,
    "args": {
      "Python parent id": 12251, "Python id": 12252, "finished": true, "Ev Idx": 15152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916376.499, "dur": 0.301,
    "args": {
      "Python parent id": 12252, "Python id": 12253, "finished": true, "Ev Idx": 15153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916376.612, "dur": 0.148,
    "args": {
      "Python parent id": 12253, "Python id": 12254, "finished": true, "Ev Idx": 15154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916377.185, "dur": 8.982,
    "args": {
      "Python parent id": 12251, "Python id": 12255, "finished": true, "Ev Idx": 15155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916377.273, "dur": 0.075,
    "args": {
      "Python parent id": 12255, "Python id": 12256, "finished": true, "Ev Idx": 15156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916377.483, "dur": 0.061,
    "args": {
      "Python parent id": 12255, "Python id": 12257, "finished": true, "Ev Idx": 15157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916377.728, "dur": 0.062,
    "args": {
      "Python parent id": 12255, "Python id": 12258, "finished": true, "Ev Idx": 15158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916378.265, "dur": 0.074,
    "args": {
      "Python parent id": 12255, "Python id": 12259, "finished": true, "Ev Idx": 15159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916378.622, "dur": 0.087,
    "args": {
      "Python parent id": 12255, "Python id": 12260, "finished": true, "Ev Idx": 15160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916379.055, "dur": 7.047,
    "args": {
      "Python parent id": 12255, "Python id": 12261, "finished": true, "Ev Idx": 15161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916379.164, "dur": 0.071,
    "args": {
      "Python parent id": 12261, "Python id": 12262, "finished": true, "Ev Idx": 15162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916379.390, "dur": 0.065,
    "args": {
      "Python parent id": 12261, "Python id": 12263, "finished": true, "Ev Idx": 15163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916379.651, "dur": 0.076,
    "args": {
      "Python parent id": 12261, "Python id": 12264, "finished": true, "Ev Idx": 15164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916380.098, "dur": 0.064,
    "args": {
      "Python parent id": 12261, "Python id": 12265, "finished": true, "Ev Idx": 15165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916380.496, "dur": 5.508,
    "args": {
      "Python parent id": 12261, "Python id": 12266, "finished": true, "Ev Idx": 15166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916380.814, "dur": 5.108,
    "args": {
      "Python parent id": 12266, "Python id": 12267, "finished": true, "Ev Idx": 15167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916380.948, "dur": 3.158,
    "args": {
      "Python parent id": 12267, "Python id": 12268, "finished": true, "Ev Idx": 15168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916381.153, "dur": 2.905,
    "args": {
      "Python parent id": 12268, "Python id": 12269, "finished": true, "Ev Idx": 15169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916381.288, "dur": 0.275,
    "args": {
      "Python parent id": 12269, "Python id": 12270, "finished": true, "Ev Idx": 15170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916381.441, "dur": 0.064,
    "args": {
      "Python parent id": 12270, "Python id": 12271, "finished": true, "Ev Idx": 15171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916381.764, "dur": 2.015,
    "args": {
      "Python parent id": 12269, "Python id": 12272, "finished": true, "Ev Idx": 15172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916381.950, "dur": 1.736,
    "args": {
      "Python parent id": 12272, "Python id": 12273, "finished": true, "Ev Idx": 15173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916382.147, "dur": 1.489,
    "args": {
      "Python parent id": 12273, "Python id": 12274, "finished": true, "Ev Idx": 15174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916382.271, "dur": 1.119,
    "args": {
      "Python parent id": 12274, "Python id": 12275, "finished": true, "Ev Idx": 15175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916382.452, "dur": 0.490,
    "args": {
      "Python parent id": 12275, "Python id": 12276, "finished": true, "Ev Idx": 15176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916382.553, "dur": 0.057,
    "args": {
      "Python parent id": 12276, "Python id": 12277, "finished": true, "Ev Idx": 15177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916382.776, "dur": 0.111,
    "args": {
      "Python parent id": 12276, "Python id": 12278, "finished": true, "Ev Idx": 15178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916383.951, "dur": 0.070,
    "args": {
      "Python parent id": 12269, "Python id": 12279, "finished": true, "Ev Idx": 15179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916384.270, "dur": 0.103,
    "args": {
      "Python parent id": 12267, "Python id": 12280, "finished": true, "Ev Idx": 15180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916384.651, "dur": 1.225,
    "args": {
      "Python parent id": 12267, "Python id": 12281, "finished": true, "Ev Idx": 15181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916384.812, "dur": 1.023,
    "args": {
      "Python parent id": 12281, "Python id": 12282, "finished": true, "Ev Idx": 15182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916384.957, "dur": 0.387,
    "args": {
      "Python parent id": 12282, "Python id": 12283, "finished": true, "Ev Idx": 15183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916385.100, "dur": 0.187,
    "args": {
      "Python parent id": 12283, "Python id": 12284, "finished": true, "Ev Idx": 15184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916385.194, "dur": 0.061,
    "args": {
      "Python parent id": 12284, "Python id": 12285, "finished": true, "Ev Idx": 15185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916385.479, "dur": 0.322,
    "args": {
      "Python parent id": 12282, "Python id": 12286, "finished": true, "Ev Idx": 15186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916386.295, "dur": 0.165,
    "args": {
      "Python parent id": 12251, "Python id": 12287, "finished": true, "Ev Idx": 15187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916387.038, "dur": 1.005,
    "args": {
      "Python parent id": 12251, "Python id": 12288, "finished": true, "Ev Idx": 15188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916387.482, "dur": 0.508,
    "args": {
      "Python parent id": 12288, "Python id": 12289, "finished": true, "Ev Idx": 15189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916388.479, "dur": 3.837,
    "args": {
      "Python parent id": 12250, "Python id": 12290, "finished": true, "Ev Idx": 15190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916393.367, "dur": 113.611,
    "args": {
      "Python parent id": 12, "Python id": 12291, "finished": true, "Ev Idx": 15191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916393.832, "dur": 0.499,
    "args": {
      "Python parent id": 12291, "Python id": 12292, "finished": true, "Ev Idx": 15192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916394.175, "dur": 0.091,
    "args": {
      "Python parent id": 12292, "Python id": 12293, "finished": true, "Ev Idx": 15193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916394.700, "dur": 0.351,
    "args": {
      "Python parent id": 12291, "Python id": 12294, "finished": true, "Ev Idx": 15194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916395.268, "dur": 0.254,
    "args": {
      "Python parent id": 12291, "Python id": 12295, "finished": true, "Ev Idx": 15195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916395.780, "dur": 1.628,
    "args": {
      "Python parent id": 12291, "Python id": 12296, "finished": true, "Ev Idx": 15196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916398.079, "dur": 108.801,
    "args": {
      "Python parent id": 12291, "Python id": 12297, "finished": true, "Ev Idx": 15197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916398.955, "dur": 107.645,
    "args": {
      "Python parent id": 12297, "Python id": 12298, "finished": true, "Ev Idx": 15198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916399.308, "dur": 107.186,
    "args": {
      "Python parent id": 12298, "Python id": 12299, "finished": true, "Ev Idx": 15199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916404.731, "dur": 99.888,
    "args": {
      "Python parent id": 12299, "Python id": 12300, "finished": true, "Ev Idx": 15200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916405.084, "dur": 99.450,
    "args": {
      "Python parent id": 12300, "Python id": 12301, "finished": true, "Ev Idx": 15201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916405.671, "dur": 98.701,
    "args": {
      "Python parent id": 12301, "Python id": 12302, "finished": true, "Ev Idx": 15202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916406.228, "dur": 0.512,
    "args": {
      "Python parent id": 12302, "Python id": 12303, "finished": true, "Ev Idx": 15203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916406.573, "dur": 0.101,
    "args": {
      "Python parent id": 12303, "Python id": 12304, "finished": true, "Ev Idx": 15204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916407.085, "dur": 0.703,
    "args": {
      "Python parent id": 12302, "Python id": 12305, "finished": true, "Ev Idx": 15205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916407.359, "dur": 0.309,
    "args": {
      "Python parent id": 12305, "Python id": 12306, "finished": true, "Ev Idx": 15206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916408.151, "dur": 96.111,
    "args": {
      "Python parent id": 12302, "Python id": 12307, "finished": true, "Ev Idx": 15207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916408.823, "dur": 93.783,
    "args": {
      "Python parent id": 12307, "Python id": 12308, "finished": true, "Ev Idx": 15208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916416.200, "dur": 18.039,
    "args": {
      "Python parent id": 12308, "Python id": 12309, "finished": true, "Ev Idx": 15209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916416.331, "dur": 0.140,
    "args": {
      "Python parent id": 12309, "Python id": 12310, "finished": true, "Ev Idx": 15210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916416.759, "dur": 17.394,
    "args": {
      "Python parent id": 12309, "Python id": 12311, "finished": true, "Ev Idx": 15211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916416.891, "dur": 7.891,
    "args": {
      "Python parent id": 12311, "Python id": 12312, "finished": true, "Ev Idx": 15212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916442.060, "dur": 6.681,
    "args": {
      "Python parent id": 12308, "Python id": 12313, "finished": true, "Ev Idx": 15213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916460.128, "dur": 20.006,
    "args": {
      "Python parent id": 12308, "Python id": 12314, "finished": true, "Ev Idx": 15214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916508.678, "dur": 21.831,
    "args": {
      "Python parent id": 12, "Python id": 12315, "finished": true, "Ev Idx": 15215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916509.116, "dur": 16.424,
    "args": {
      "Python parent id": 12315, "Python id": 12316, "finished": true, "Ev Idx": 15216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916509.312, "dur": 0.668,
    "args": {
      "Python parent id": 12316, "Python id": 12317, "finished": true, "Ev Idx": 15217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916509.494, "dur": 0.384,
    "args": {
      "Python parent id": 12317, "Python id": 12318, "finished": true, "Ev Idx": 15218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916509.689, "dur": 0.139,
    "args": {
      "Python parent id": 12318, "Python id": 12319, "finished": true, "Ev Idx": 15219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916510.536, "dur": 11.832,
    "args": {
      "Python parent id": 12316, "Python id": 12320, "finished": true, "Ev Idx": 15220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916510.694, "dur": 0.098,
    "args": {
      "Python parent id": 12320, "Python id": 12321, "finished": true, "Ev Idx": 15221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916510.934, "dur": 0.066,
    "args": {
      "Python parent id": 12320, "Python id": 12322, "finished": true, "Ev Idx": 15222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916511.174, "dur": 0.059,
    "args": {
      "Python parent id": 12320, "Python id": 12323, "finished": true, "Ev Idx": 15223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916511.711, "dur": 0.099,
    "args": {
      "Python parent id": 12320, "Python id": 12324, "finished": true, "Ev Idx": 15224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916512.138, "dur": 0.091,
    "args": {
      "Python parent id": 12320, "Python id": 12325, "finished": true, "Ev Idx": 15225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916512.664, "dur": 9.618,
    "args": {
      "Python parent id": 12320, "Python id": 12326, "finished": true, "Ev Idx": 15226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916512.748, "dur": 0.061,
    "args": {
      "Python parent id": 12326, "Python id": 12327, "finished": true, "Ev Idx": 15227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916512.990, "dur": 0.062,
    "args": {
      "Python parent id": 12326, "Python id": 12328, "finished": true, "Ev Idx": 15228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916513.233, "dur": 0.081,
    "args": {
      "Python parent id": 12326, "Python id": 12329, "finished": true, "Ev Idx": 15229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916513.676, "dur": 0.082,
    "args": {
      "Python parent id": 12326, "Python id": 12330, "finished": true, "Ev Idx": 15230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916514.146, "dur": 7.980,
    "args": {
      "Python parent id": 12326, "Python id": 12331, "finished": true, "Ev Idx": 15231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916514.476, "dur": 7.589,
    "args": {
      "Python parent id": 12331, "Python id": 12332, "finished": true, "Ev Idx": 15232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916514.649, "dur": 4.376,
    "args": {
      "Python parent id": 12332, "Python id": 12333, "finished": true, "Ev Idx": 15233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916514.901, "dur": 4.030,
    "args": {
      "Python parent id": 12333, "Python id": 12334, "finished": true, "Ev Idx": 15234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916515.067, "dur": 0.302,
    "args": {
      "Python parent id": 12334, "Python id": 12335, "finished": true, "Ev Idx": 15235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916515.219, "dur": 0.064,
    "args": {
      "Python parent id": 12335, "Python id": 12336, "finished": true, "Ev Idx": 15236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916515.558, "dur": 2.924,
    "args": {
      "Python parent id": 12334, "Python id": 12337, "finished": true, "Ev Idx": 15237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916515.800, "dur": 2.553,
    "args": {
      "Python parent id": 12337, "Python id": 12338, "finished": true, "Ev Idx": 15238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916516.059, "dur": 2.209,
    "args": {
      "Python parent id": 12338, "Python id": 12339, "finished": true, "Ev Idx": 15239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916516.211, "dur": 1.709,
    "args": {
      "Python parent id": 12339, "Python id": 12340, "finished": true, "Ev Idx": 15240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916516.469, "dur": 0.632,
    "args": {
      "Python parent id": 12340, "Python id": 12341, "finished": true, "Ev Idx": 15241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916516.575, "dur": 0.049,
    "args": {
      "Python parent id": 12341, "Python id": 12342, "finished": true, "Ev Idx": 15242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916516.838, "dur": 0.189,
    "args": {
      "Python parent id": 12341, "Python id": 12343, "finished": true, "Ev Idx": 15243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916518.666, "dur": 0.205,
    "args": {
      "Python parent id": 12334, "Python id": 12344, "finished": true, "Ev Idx": 15244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916520.084, "dur": 0.122,
    "args": {
      "Python parent id": 12332, "Python id": 12345, "finished": true, "Ev Idx": 15245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916520.585, "dur": 1.436,
    "args": {
      "Python parent id": 12332, "Python id": 12346, "finished": true, "Ev Idx": 15246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916520.804, "dur": 1.157,
    "args": {
      "Python parent id": 12346, "Python id": 12347, "finished": true, "Ev Idx": 15247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916521.005, "dur": 0.451,
    "args": {
      "Python parent id": 12347, "Python id": 12348, "finished": true, "Ev Idx": 15248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916521.156, "dur": 0.225,
    "args": {
      "Python parent id": 12348, "Python id": 12349, "finished": true, "Ev Idx": 15249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916521.274, "dur": 0.055,
    "args": {
      "Python parent id": 12349, "Python id": 12350, "finished": true, "Ev Idx": 15250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916521.602, "dur": 0.312,
    "args": {
      "Python parent id": 12347, "Python id": 12351, "finished": true, "Ev Idx": 15251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916522.566, "dur": 0.360,
    "args": {
      "Python parent id": 12316, "Python id": 12352, "finished": true, "Ev Idx": 15252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916523.807, "dur": 1.544,
    "args": {
      "Python parent id": 12316, "Python id": 12353, "finished": true, "Ev Idx": 15253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916524.474, "dur": 0.784,
    "args": {
      "Python parent id": 12353, "Python id": 12354, "finished": true, "Ev Idx": 15254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916526.001, "dur": 4.409,
    "args": {
      "Python parent id": 12315, "Python id": 12355, "finished": true, "Ev Idx": 15255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916531.153, "dur": 8.373,
    "args": {
      "Python parent id": 12, "Python id": 12356, "finished": true, "Ev Idx": 15256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916540.022, "dur": 0.147,
    "args": {
      "Python parent id": 12, "Python id": 12357, "finished": true, "Ev Idx": 15257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916540.923, "dur": 0.266,
    "args": {
      "Python parent id": 12, "Python id": 12358, "finished": true, "Ev Idx": 15258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916541.736, "dur": 0.072,
    "args": {
      "Python parent id": 12, "Python id": 12359, "finished": true, "Ev Idx": 15259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916542.210, "dur": 11.387,
    "args": {
      "Python parent id": 12, "Python id": 12360, "finished": true, "Ev Idx": 15260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916555.110, "dur": 17.049,
    "args": {
      "Python parent id": 12, "Python id": 12361, "finished": true, "Ev Idx": 15261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916555.491, "dur": 12.808,
    "args": {
      "Python parent id": 12361, "Python id": 12362, "finished": true, "Ev Idx": 15262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916555.690, "dur": 0.602,
    "args": {
      "Python parent id": 12362, "Python id": 12363, "finished": true, "Ev Idx": 15263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916555.858, "dur": 0.327,
    "args": {
      "Python parent id": 12363, "Python id": 12364, "finished": true, "Ev Idx": 15264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916555.995, "dur": 0.118,
    "args": {
      "Python parent id": 12364, "Python id": 12365, "finished": true, "Ev Idx": 15265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916556.568, "dur": 9.531,
    "args": {
      "Python parent id": 12362, "Python id": 12366, "finished": true, "Ev Idx": 15266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916556.656, "dur": 0.070,
    "args": {
      "Python parent id": 12366, "Python id": 12367, "finished": true, "Ev Idx": 15267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916556.839, "dur": 0.067,
    "args": {
      "Python parent id": 12366, "Python id": 12368, "finished": true, "Ev Idx": 15268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916557.102, "dur": 0.065,
    "args": {
      "Python parent id": 12366, "Python id": 12369, "finished": true, "Ev Idx": 15269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916557.629, "dur": 0.068,
    "args": {
      "Python parent id": 12366, "Python id": 12370, "finished": true, "Ev Idx": 15270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916557.974, "dur": 0.079,
    "args": {
      "Python parent id": 12366, "Python id": 12371, "finished": true, "Ev Idx": 15271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916558.480, "dur": 7.566,
    "args": {
      "Python parent id": 12366, "Python id": 12372, "finished": true, "Ev Idx": 15272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916558.574, "dur": 0.066,
    "args": {
      "Python parent id": 12372, "Python id": 12373, "finished": true, "Ev Idx": 15273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916558.812, "dur": 0.056,
    "args": {
      "Python parent id": 12372, "Python id": 12374, "finished": true, "Ev Idx": 15274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916559.049, "dur": 0.069,
    "args": {
      "Python parent id": 12372, "Python id": 12375, "finished": true, "Ev Idx": 15275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916559.574, "dur": 0.038,
    "args": {
      "Python parent id": 12372, "Python id": 12376, "finished": true, "Ev Idx": 15276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916559.934, "dur": 5.990,
    "args": {
      "Python parent id": 12372, "Python id": 12377, "finished": true, "Ev Idx": 15277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916560.152, "dur": 5.693,
    "args": {
      "Python parent id": 12377, "Python id": 12378, "finished": true, "Ev Idx": 15278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916560.298, "dur": 3.556,
    "args": {
      "Python parent id": 12378, "Python id": 12379, "finished": true, "Ev Idx": 15279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916560.549, "dur": 3.209,
    "args": {
      "Python parent id": 12379, "Python id": 12380, "finished": true, "Ev Idx": 15280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916560.679, "dur": 0.314,
    "args": {
      "Python parent id": 12380, "Python id": 12381, "finished": true, "Ev Idx": 15281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916560.826, "dur": 0.064,
    "args": {
      "Python parent id": 12381, "Python id": 12382, "finished": true, "Ev Idx": 15282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916561.217, "dur": 2.221,
    "args": {
      "Python parent id": 12380, "Python id": 12383, "finished": true, "Ev Idx": 15283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916561.381, "dur": 1.970,
    "args": {
      "Python parent id": 12383, "Python id": 12384, "finished": true, "Ev Idx": 15284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916561.618, "dur": 1.667,
    "args": {
      "Python parent id": 12384, "Python id": 12385, "finished": true, "Ev Idx": 15285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916561.765, "dur": 1.288,
    "args": {
      "Python parent id": 12385, "Python id": 12386, "finished": true, "Ev Idx": 15286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916561.973, "dur": 0.563,
    "args": {
      "Python parent id": 12386, "Python id": 12387, "finished": true, "Ev Idx": 15287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916562.072, "dur": 0.053,
    "args": {
      "Python parent id": 12387, "Python id": 12388, "finished": true, "Ev Idx": 15288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916562.291, "dur": 0.162,
    "args": {
      "Python parent id": 12387, "Python id": 12389, "finished": true, "Ev Idx": 15289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916563.610, "dur": 0.111,
    "args": {
      "Python parent id": 12380, "Python id": 12390, "finished": true, "Ev Idx": 15290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916564.050, "dur": 0.122,
    "args": {
      "Python parent id": 12378, "Python id": 12391, "finished": true, "Ev Idx": 15291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916564.508, "dur": 1.273,
    "args": {
      "Python parent id": 12378, "Python id": 12392, "finished": true, "Ev Idx": 15292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916564.682, "dur": 1.035,
    "args": {
      "Python parent id": 12392, "Python id": 12393, "finished": true, "Ev Idx": 15293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916564.821, "dur": 0.454,
    "args": {
      "Python parent id": 12393, "Python id": 12394, "finished": true, "Ev Idx": 15294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916564.965, "dur": 0.204,
    "args": {
      "Python parent id": 12394, "Python id": 12395, "finished": true, "Ev Idx": 15295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916565.078, "dur": 0.060,
    "args": {
      "Python parent id": 12395, "Python id": 12396, "finished": true, "Ev Idx": 15296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916565.420, "dur": 0.269,
    "args": {
      "Python parent id": 12393, "Python id": 12397, "finished": true, "Ev Idx": 15297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916566.244, "dur": 0.189,
    "args": {
      "Python parent id": 12362, "Python id": 12398, "finished": true, "Ev Idx": 15298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916567.085, "dur": 1.043,
    "args": {
      "Python parent id": 12362, "Python id": 12399, "finished": true, "Ev Idx": 15299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916567.524, "dur": 0.536,
    "args": {
      "Python parent id": 12399, "Python id": 12400, "finished": true, "Ev Idx": 15300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916568.590, "dur": 3.494,
    "args": {
      "Python parent id": 12361, "Python id": 12401, "finished": true, "Ev Idx": 15301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916573.122, "dur": 98.519,
    "args": {
      "Python parent id": 12, "Python id": 12402, "finished": true, "Ev Idx": 15302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916573.579, "dur": 0.541,
    "args": {
      "Python parent id": 12402, "Python id": 12403, "finished": true, "Ev Idx": 15303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916573.964, "dur": 0.088,
    "args": {
      "Python parent id": 12403, "Python id": 12404, "finished": true, "Ev Idx": 15304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916574.500, "dur": 0.559,
    "args": {
      "Python parent id": 12402, "Python id": 12405, "finished": true, "Ev Idx": 15305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916575.290, "dur": 0.263,
    "args": {
      "Python parent id": 12402, "Python id": 12406, "finished": true, "Ev Idx": 15306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916575.806, "dur": 1.846,
    "args": {
      "Python parent id": 12402, "Python id": 12407, "finished": true, "Ev Idx": 15307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916578.372, "dur": 93.139,
    "args": {
      "Python parent id": 12402, "Python id": 12408, "finished": true, "Ev Idx": 15308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916579.330, "dur": 91.963,
    "args": {
      "Python parent id": 12408, "Python id": 12409, "finished": true, "Ev Idx": 15309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916579.717, "dur": 91.493,
    "args": {
      "Python parent id": 12409, "Python id": 12410, "finished": true, "Ev Idx": 15310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916586.458, "dur": 83.241,
    "args": {
      "Python parent id": 12410, "Python id": 12411, "finished": true, "Ev Idx": 15311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916586.801, "dur": 82.844,
    "args": {
      "Python parent id": 12411, "Python id": 12412, "finished": true, "Ev Idx": 15312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916587.469, "dur": 81.993,
    "args": {
      "Python parent id": 12412, "Python id": 12413, "finished": true, "Ev Idx": 15313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916588.049, "dur": 0.494,
    "args": {
      "Python parent id": 12413, "Python id": 12414, "finished": true, "Ev Idx": 15314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916588.377, "dur": 0.124,
    "args": {
      "Python parent id": 12414, "Python id": 12415, "finished": true, "Ev Idx": 15315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916588.881, "dur": 0.388,
    "args": {
      "Python parent id": 12413, "Python id": 12416, "finished": true, "Ev Idx": 15316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916589.162, "dur": 0.052,
    "args": {
      "Python parent id": 12416, "Python id": 12417, "finished": true, "Ev Idx": 15317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916589.653, "dur": 79.640,
    "args": {
      "Python parent id": 12413, "Python id": 12418, "finished": true, "Ev Idx": 15318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916590.489, "dur": 77.297,
    "args": {
      "Python parent id": 12418, "Python id": 12419, "finished": true, "Ev Idx": 15319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916599.029, "dur": 19.242,
    "args": {
      "Python parent id": 12419, "Python id": 12420, "finished": true, "Ev Idx": 15320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916599.205, "dur": 0.137,
    "args": {
      "Python parent id": 12420, "Python id": 12421, "finished": true, "Ev Idx": 15321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916599.747, "dur": 18.407,
    "args": {
      "Python parent id": 12420, "Python id": 12422, "finished": true, "Ev Idx": 15322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916599.881, "dur": 8.278,
    "args": {
      "Python parent id": 12422, "Python id": 12423, "finished": true, "Ev Idx": 15323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916627.050, "dur": 7.092,
    "args": {
      "Python parent id": 12419, "Python id": 12424, "finished": true, "Ev Idx": 15324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916643.478, "dur": 8.494,
    "args": {
      "Python parent id": 12419, "Python id": 12425, "finished": true, "Ev Idx": 15325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916673.115, "dur": 18.276,
    "args": {
      "Python parent id": 12, "Python id": 12426, "finished": true, "Ev Idx": 15326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916673.522, "dur": 14.039,
    "args": {
      "Python parent id": 12426, "Python id": 12427, "finished": true, "Ev Idx": 15327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916673.716, "dur": 0.584,
    "args": {
      "Python parent id": 12427, "Python id": 12428, "finished": true, "Ev Idx": 15328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916673.916, "dur": 0.316,
    "args": {
      "Python parent id": 12428, "Python id": 12429, "finished": true, "Ev Idx": 15329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916674.055, "dur": 0.142,
    "args": {
      "Python parent id": 12429, "Python id": 12430, "finished": true, "Ev Idx": 15330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916674.753, "dur": 10.018,
    "args": {
      "Python parent id": 12427, "Python id": 12431, "finished": true, "Ev Idx": 15331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916674.908, "dur": 0.063,
    "args": {
      "Python parent id": 12431, "Python id": 12432, "finished": true, "Ev Idx": 15332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916675.126, "dur": 0.060,
    "args": {
      "Python parent id": 12431, "Python id": 12433, "finished": true, "Ev Idx": 15333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916675.379, "dur": 0.059,
    "args": {
      "Python parent id": 12431, "Python id": 12434, "finished": true, "Ev Idx": 15334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916675.913, "dur": 0.066,
    "args": {
      "Python parent id": 12431, "Python id": 12435, "finished": true, "Ev Idx": 15335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916676.292, "dur": 0.094,
    "args": {
      "Python parent id": 12431, "Python id": 12436, "finished": true, "Ev Idx": 15336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916676.743, "dur": 7.977,
    "args": {
      "Python parent id": 12431, "Python id": 12437, "finished": true, "Ev Idx": 15337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916676.834, "dur": 0.067,
    "args": {
      "Python parent id": 12437, "Python id": 12438, "finished": true, "Ev Idx": 15338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916677.064, "dur": 0.054,
    "args": {
      "Python parent id": 12437, "Python id": 12439, "finished": true, "Ev Idx": 15339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916677.294, "dur": 0.069,
    "args": {
      "Python parent id": 12437, "Python id": 12440, "finished": true, "Ev Idx": 15340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916677.762, "dur": 0.054,
    "args": {
      "Python parent id": 12437, "Python id": 12441, "finished": true, "Ev Idx": 15341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916678.118, "dur": 6.492,
    "args": {
      "Python parent id": 12437, "Python id": 12442, "finished": true, "Ev Idx": 15342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916678.430, "dur": 6.095,
    "args": {
      "Python parent id": 12442, "Python id": 12443, "finished": true, "Ev Idx": 15343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916678.559, "dur": 3.968,
    "args": {
      "Python parent id": 12443, "Python id": 12444, "finished": true, "Ev Idx": 15344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916678.837, "dur": 3.616,
    "args": {
      "Python parent id": 12444, "Python id": 12445, "finished": true, "Ev Idx": 15345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916678.951, "dur": 0.309,
    "args": {
      "Python parent id": 12445, "Python id": 12446, "finished": true, "Ev Idx": 15346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916679.136, "dur": 0.058,
    "args": {
      "Python parent id": 12446, "Python id": 12447, "finished": true, "Ev Idx": 15347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916679.453, "dur": 2.606,
    "args": {
      "Python parent id": 12445, "Python id": 12448, "finished": true, "Ev Idx": 15348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916679.694, "dur": 2.223,
    "args": {
      "Python parent id": 12448, "Python id": 12449, "finished": true, "Ev Idx": 15349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916679.966, "dur": 1.866,
    "args": {
      "Python parent id": 12449, "Python id": 12450, "finished": true, "Ev Idx": 15350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916680.167, "dur": 1.436,
    "args": {
      "Python parent id": 12450, "Python id": 12451, "finished": true, "Ev Idx": 15351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916680.379, "dur": 0.700,
    "args": {
      "Python parent id": 12451, "Python id": 12452, "finished": true, "Ev Idx": 15352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916680.492, "dur": 0.065,
    "args": {
      "Python parent id": 12452, "Python id": 12453, "finished": true, "Ev Idx": 15353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916680.797, "dur": 0.201,
    "args": {
      "Python parent id": 12452, "Python id": 12454, "finished": true, "Ev Idx": 15354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916682.256, "dur": 0.164,
    "args": {
      "Python parent id": 12445, "Python id": 12455, "finished": true, "Ev Idx": 15355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916682.754, "dur": 0.119,
    "args": {
      "Python parent id": 12443, "Python id": 12456, "finished": true, "Ev Idx": 15356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916683.195, "dur": 1.292,
    "args": {
      "Python parent id": 12443, "Python id": 12457, "finished": true, "Ev Idx": 15357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916683.364, "dur": 1.047,
    "args": {
      "Python parent id": 12457, "Python id": 12458, "finished": true, "Ev Idx": 15358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916683.503, "dur": 0.439,
    "args": {
      "Python parent id": 12458, "Python id": 12459, "finished": true, "Ev Idx": 15359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916683.660, "dur": 0.208,
    "args": {
      "Python parent id": 12459, "Python id": 12460, "finished": true, "Ev Idx": 15360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916683.757, "dur": 0.080,
    "args": {
      "Python parent id": 12460, "Python id": 12461, "finished": true, "Ev Idx": 15361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916684.096, "dur": 0.260,
    "args": {
      "Python parent id": 12458, "Python id": 12462, "finished": true, "Ev Idx": 15362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916684.921, "dur": 0.413,
    "args": {
      "Python parent id": 12427, "Python id": 12463, "finished": true, "Ev Idx": 15363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916686.025, "dur": 1.370,
    "args": {
      "Python parent id": 12427, "Python id": 12464, "finished": true, "Ev Idx": 15364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916686.618, "dur": 0.707,
    "args": {
      "Python parent id": 12464, "Python id": 12465, "finished": true, "Ev Idx": 15365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916687.913, "dur": 3.383,
    "args": {
      "Python parent id": 12426, "Python id": 12466, "finished": true, "Ev Idx": 15366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916692.006, "dur": 9.520,
    "args": {
      "Python parent id": 12, "Python id": 12467, "finished": true, "Ev Idx": 15367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916702.024, "dur": 0.159,
    "args": {
      "Python parent id": 12, "Python id": 12468, "finished": true, "Ev Idx": 15368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916703.002, "dur": 0.230,
    "args": {
      "Python parent id": 12, "Python id": 12469, "finished": true, "Ev Idx": 15369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916703.767, "dur": 0.080,
    "args": {
      "Python parent id": 12, "Python id": 12470, "finished": true, "Ev Idx": 15370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916704.254, "dur": 10.735,
    "args": {
      "Python parent id": 12, "Python id": 12471, "finished": true, "Ev Idx": 15371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916716.319, "dur": 16.188,
    "args": {
      "Python parent id": 12, "Python id": 12472, "finished": true, "Ev Idx": 15372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916716.672, "dur": 12.290,
    "args": {
      "Python parent id": 12472, "Python id": 12473, "finished": true, "Ev Idx": 15373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916716.820, "dur": 0.583,
    "args": {
      "Python parent id": 12473, "Python id": 12474, "finished": true, "Ev Idx": 15374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916717.001, "dur": 0.255,
    "args": {
      "Python parent id": 12474, "Python id": 12475, "finished": true, "Ev Idx": 15375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916717.119, "dur": 0.097,
    "args": {
      "Python parent id": 12475, "Python id": 12476, "finished": true, "Ev Idx": 15376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916717.726, "dur": 9.121,
    "args": {
      "Python parent id": 12473, "Python id": 12477, "finished": true, "Ev Idx": 15377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916717.812, "dur": 0.075,
    "args": {
      "Python parent id": 12477, "Python id": 12478, "finished": true, "Ev Idx": 15378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916718.045, "dur": 0.062,
    "args": {
      "Python parent id": 12477, "Python id": 12479, "finished": true, "Ev Idx": 15379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916718.276, "dur": 0.059,
    "args": {
      "Python parent id": 12477, "Python id": 12480, "finished": true, "Ev Idx": 15380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916718.752, "dur": 0.097,
    "args": {
      "Python parent id": 12477, "Python id": 12481, "finished": true, "Ev Idx": 15381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916719.134, "dur": 0.081,
    "args": {
      "Python parent id": 12477, "Python id": 12482, "finished": true, "Ev Idx": 15382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916719.569, "dur": 7.201,
    "args": {
      "Python parent id": 12477, "Python id": 12483, "finished": true, "Ev Idx": 15383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916719.668, "dur": 0.066,
    "args": {
      "Python parent id": 12483, "Python id": 12484, "finished": true, "Ev Idx": 15384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916719.894, "dur": 0.064,
    "args": {
      "Python parent id": 12483, "Python id": 12485, "finished": true, "Ev Idx": 15385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916720.154, "dur": 0.071,
    "args": {
      "Python parent id": 12483, "Python id": 12486, "finished": true, "Ev Idx": 15386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916720.581, "dur": 0.038,
    "args": {
      "Python parent id": 12483, "Python id": 12487, "finished": true, "Ev Idx": 15387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916720.919, "dur": 5.776,
    "args": {
      "Python parent id": 12483, "Python id": 12488, "finished": true, "Ev Idx": 15388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916721.224, "dur": 5.375,
    "args": {
      "Python parent id": 12488, "Python id": 12489, "finished": true, "Ev Idx": 15389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916721.363, "dur": 3.417,
    "args": {
      "Python parent id": 12489, "Python id": 12490, "finished": true, "Ev Idx": 15390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916721.628, "dur": 3.071,
    "args": {
      "Python parent id": 12490, "Python id": 12491, "finished": true, "Ev Idx": 15391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916721.752, "dur": 0.276,
    "args": {
      "Python parent id": 12491, "Python id": 12492, "finished": true, "Ev Idx": 15392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916721.907, "dur": 0.067,
    "args": {
      "Python parent id": 12492, "Python id": 12493, "finished": true, "Ev Idx": 15393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916722.240, "dur": 2.142,
    "args": {
      "Python parent id": 12491, "Python id": 12494, "finished": true, "Ev Idx": 15394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916722.466, "dur": 1.829,
    "args": {
      "Python parent id": 12494, "Python id": 12495, "finished": true, "Ev Idx": 15395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916722.665, "dur": 1.579,
    "args": {
      "Python parent id": 12495, "Python id": 12496, "finished": true, "Ev Idx": 15396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916722.824, "dur": 1.214,
    "args": {
      "Python parent id": 12496, "Python id": 12497, "finished": true, "Ev Idx": 15397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916722.996, "dur": 0.548,
    "args": {
      "Python parent id": 12497, "Python id": 12498, "finished": true, "Ev Idx": 15398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916723.080, "dur": 0.077,
    "args": {
      "Python parent id": 12498, "Python id": 12499, "finished": true, "Ev Idx": 15399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916723.333, "dur": 0.167,
    "args": {
      "Python parent id": 12498, "Python id": 12500, "finished": true, "Ev Idx": 15400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916724.570, "dur": 0.096,
    "args": {
      "Python parent id": 12491, "Python id": 12501, "finished": true, "Ev Idx": 15401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916724.982, "dur": 0.130,
    "args": {
      "Python parent id": 12489, "Python id": 12502, "finished": true, "Ev Idx": 15402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916725.403, "dur": 1.153,
    "args": {
      "Python parent id": 12489, "Python id": 12503, "finished": true, "Ev Idx": 15403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916725.541, "dur": 0.972,
    "args": {
      "Python parent id": 12503, "Python id": 12504, "finished": true, "Ev Idx": 15404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916725.688, "dur": 0.381,
    "args": {
      "Python parent id": 12504, "Python id": 12505, "finished": true, "Ev Idx": 15405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916725.827, "dur": 0.181,
    "args": {
      "Python parent id": 12505, "Python id": 12506, "finished": true, "Ev Idx": 15406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916725.921, "dur": 0.051,
    "args": {
      "Python parent id": 12506, "Python id": 12507, "finished": true, "Ev Idx": 15407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916726.208, "dur": 0.276,
    "args": {
      "Python parent id": 12504, "Python id": 12508, "finished": true, "Ev Idx": 15408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916726.972, "dur": 0.180,
    "args": {
      "Python parent id": 12473, "Python id": 12509, "finished": true, "Ev Idx": 15409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916727.795, "dur": 0.982,
    "args": {
      "Python parent id": 12473, "Python id": 12510, "finished": true, "Ev Idx": 15410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916728.208, "dur": 0.515,
    "args": {
      "Python parent id": 12510, "Python id": 12511, "finished": true, "Ev Idx": 15411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916729.295, "dur": 3.108,
    "args": {
      "Python parent id": 12472, "Python id": 12512, "finished": true, "Ev Idx": 15412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916733.543, "dur": 94.105,
    "args": {
      "Python parent id": 12, "Python id": 12513, "finished": true, "Ev Idx": 15413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916733.991, "dur": 0.471,
    "args": {
      "Python parent id": 12513, "Python id": 12514, "finished": true, "Ev Idx": 15414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916734.262, "dur": 0.118,
    "args": {
      "Python parent id": 12514, "Python id": 12515, "finished": true, "Ev Idx": 15415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916734.811, "dur": 0.434,
    "args": {
      "Python parent id": 12513, "Python id": 12516, "finished": true, "Ev Idx": 15416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916735.463, "dur": 0.290,
    "args": {
      "Python parent id": 12513, "Python id": 12517, "finished": true, "Ev Idx": 15417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916735.990, "dur": 1.758,
    "args": {
      "Python parent id": 12513, "Python id": 12518, "finished": true, "Ev Idx": 15418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916738.414, "dur": 89.114,
    "args": {
      "Python parent id": 12513, "Python id": 12519, "finished": true, "Ev Idx": 15419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916739.316, "dur": 87.966,
    "args": {
      "Python parent id": 12519, "Python id": 12520, "finished": true, "Ev Idx": 15420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916739.663, "dur": 87.554,
    "args": {
      "Python parent id": 12520, "Python id": 12521, "finished": true, "Ev Idx": 15421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916745.723, "dur": 80.121,
    "args": {
      "Python parent id": 12521, "Python id": 12522, "finished": true, "Ev Idx": 15422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916746.087, "dur": 79.644,
    "args": {
      "Python parent id": 12522, "Python id": 12523, "finished": true, "Ev Idx": 15423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916746.677, "dur": 78.905,
    "args": {
      "Python parent id": 12523, "Python id": 12524, "finished": true, "Ev Idx": 15424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916747.253, "dur": 0.457,
    "args": {
      "Python parent id": 12524, "Python id": 12525, "finished": true, "Ev Idx": 15425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916747.526, "dur": 0.113,
    "args": {
      "Python parent id": 12525, "Python id": 12526, "finished": true, "Ev Idx": 15426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916748.031, "dur": 0.353,
    "args": {
      "Python parent id": 12524, "Python id": 12527, "finished": true, "Ev Idx": 15427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916748.251, "dur": 0.089,
    "args": {
      "Python parent id": 12527, "Python id": 12528, "finished": true, "Ev Idx": 15428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916748.755, "dur": 76.651,
    "args": {
      "Python parent id": 12524, "Python id": 12529, "finished": true, "Ev Idx": 15429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916749.577, "dur": 74.452,
    "args": {
      "Python parent id": 12529, "Python id": 12530, "finished": true, "Ev Idx": 15430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916757.051, "dur": 19.049,
    "args": {
      "Python parent id": 12530, "Python id": 12531, "finished": true, "Ev Idx": 15431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916757.210, "dur": 0.137,
    "args": {
      "Python parent id": 12531, "Python id": 12532, "finished": true, "Ev Idx": 15432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916757.715, "dur": 18.302,
    "args": {
      "Python parent id": 12531, "Python id": 12533, "finished": true, "Ev Idx": 15433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916757.837, "dur": 7.517,
    "args": {
      "Python parent id": 12533, "Python id": 12534, "finished": true, "Ev Idx": 15434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916784.158, "dur": 7.461,
    "args": {
      "Python parent id": 12530, "Python id": 12535, "finished": true, "Ev Idx": 15435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916800.475, "dur": 8.394,
    "args": {
      "Python parent id": 12530, "Python id": 12536, "finished": true, "Ev Idx": 15436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916829.043, "dur": 17.606,
    "args": {
      "Python parent id": 12, "Python id": 12537, "finished": true, "Ev Idx": 15437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916829.425, "dur": 12.958,
    "args": {
      "Python parent id": 12537, "Python id": 12538, "finished": true, "Ev Idx": 15438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916829.601, "dur": 0.547,
    "args": {
      "Python parent id": 12538, "Python id": 12539, "finished": true, "Ev Idx": 15439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916829.766, "dur": 0.312,
    "args": {
      "Python parent id": 12539, "Python id": 12540, "finished": true, "Ev Idx": 15440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916829.909, "dur": 0.134,
    "args": {
      "Python parent id": 12540, "Python id": 12541, "finished": true, "Ev Idx": 15441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916830.518, "dur": 9.516,
    "args": {
      "Python parent id": 12538, "Python id": 12542, "finished": true, "Ev Idx": 15442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916830.640, "dur": 0.066,
    "args": {
      "Python parent id": 12542, "Python id": 12543, "finished": true, "Ev Idx": 15443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916830.856, "dur": 0.073,
    "args": {
      "Python parent id": 12542, "Python id": 12544, "finished": true, "Ev Idx": 15444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916831.094, "dur": 0.055,
    "args": {
      "Python parent id": 12542, "Python id": 12545, "finished": true, "Ev Idx": 15445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916831.620, "dur": 0.035,
    "args": {
      "Python parent id": 12542, "Python id": 12546, "finished": true, "Ev Idx": 15446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916831.982, "dur": 0.082,
    "args": {
      "Python parent id": 12542, "Python id": 12547, "finished": true, "Ev Idx": 15447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916832.469, "dur": 7.517,
    "args": {
      "Python parent id": 12542, "Python id": 12548, "finished": true, "Ev Idx": 15448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916832.565, "dur": 0.064,
    "args": {
      "Python parent id": 12548, "Python id": 12549, "finished": true, "Ev Idx": 15449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916832.805, "dur": 0.060,
    "args": {
      "Python parent id": 12548, "Python id": 12550, "finished": true, "Ev Idx": 15450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916833.109, "dur": 0.075,
    "args": {
      "Python parent id": 12548, "Python id": 12551, "finished": true, "Ev Idx": 15451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916833.605, "dur": 0.039,
    "args": {
      "Python parent id": 12548, "Python id": 12552, "finished": true, "Ev Idx": 15452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916833.980, "dur": 5.897,
    "args": {
      "Python parent id": 12548, "Python id": 12553, "finished": true, "Ev Idx": 15453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916834.288, "dur": 5.501,
    "args": {
      "Python parent id": 12553, "Python id": 12554, "finished": true, "Ev Idx": 15454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916834.444, "dur": 3.514,
    "args": {
      "Python parent id": 12554, "Python id": 12555, "finished": true, "Ev Idx": 15455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916834.657, "dur": 3.251,
    "args": {
      "Python parent id": 12555, "Python id": 12556, "finished": true, "Ev Idx": 15456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916834.821, "dur": 0.256,
    "args": {
      "Python parent id": 12556, "Python id": 12557, "finished": true, "Ev Idx": 15457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916834.954, "dur": 0.067,
    "args": {
      "Python parent id": 12557, "Python id": 12558, "finished": true, "Ev Idx": 15458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916835.277, "dur": 2.264,
    "args": {
      "Python parent id": 12556, "Python id": 12559, "finished": true, "Ev Idx": 15459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916835.525, "dur": 1.927,
    "args": {
      "Python parent id": 12559, "Python id": 12560, "finished": true, "Ev Idx": 15460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916835.750, "dur": 1.654,
    "args": {
      "Python parent id": 12560, "Python id": 12561, "finished": true, "Ev Idx": 15461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916835.911, "dur": 1.294,
    "args": {
      "Python parent id": 12561, "Python id": 12562, "finished": true, "Ev Idx": 15462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916836.109, "dur": 0.518,
    "args": {
      "Python parent id": 12562, "Python id": 12563, "finished": true, "Ev Idx": 15463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916836.197, "dur": 0.047,
    "args": {
      "Python parent id": 12563, "Python id": 12564, "finished": true, "Ev Idx": 15464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916836.428, "dur": 0.144,
    "args": {
      "Python parent id": 12563, "Python id": 12565, "finished": true, "Ev Idx": 15465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916837.749, "dur": 0.126,
    "args": {
      "Python parent id": 12556, "Python id": 12566, "finished": true, "Ev Idx": 15466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916838.161, "dur": 0.103,
    "args": {
      "Python parent id": 12554, "Python id": 12567, "finished": true, "Ev Idx": 15467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916838.515, "dur": 1.229,
    "args": {
      "Python parent id": 12554, "Python id": 12568, "finished": true, "Ev Idx": 15468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916838.713, "dur": 0.967,
    "args": {
      "Python parent id": 12568, "Python id": 12569, "finished": true, "Ev Idx": 15469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916838.848, "dur": 0.394,
    "args": {
      "Python parent id": 12569, "Python id": 12570, "finished": true, "Ev Idx": 15470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916838.988, "dur": 0.197,
    "args": {
      "Python parent id": 12570, "Python id": 12571, "finished": true, "Ev Idx": 15471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916839.077, "dur": 0.077,
    "args": {
      "Python parent id": 12571, "Python id": 12572, "finished": true, "Ev Idx": 15472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916839.368, "dur": 0.257,
    "args": {
      "Python parent id": 12569, "Python id": 12573, "finished": true, "Ev Idx": 15473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916840.182, "dur": 0.258,
    "args": {
      "Python parent id": 12538, "Python id": 12574, "finished": true, "Ev Idx": 15474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916841.140, "dur": 1.110,
    "args": {
      "Python parent id": 12538, "Python id": 12575, "finished": true, "Ev Idx": 15475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916841.668, "dur": 0.526,
    "args": {
      "Python parent id": 12575, "Python id": 12576, "finished": true, "Ev Idx": 15476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916842.760, "dur": 3.790,
    "args": {
      "Python parent id": 12537, "Python id": 12577, "finished": true, "Ev Idx": 15477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916847.344, "dur": 7.958,
    "args": {
      "Python parent id": 12, "Python id": 12578, "finished": true, "Ev Idx": 15478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740916855.817, "dur": 0.123,
    "args": {
      "Python parent id": 12, "Python id": 12579, "finished": true, "Ev Idx": 15479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916856.617, "dur": 0.245,
    "args": {
      "Python parent id": 12, "Python id": 12580, "finished": true, "Ev Idx": 15480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916857.425, "dur": 0.073,
    "args": {
      "Python parent id": 12, "Python id": 12581, "finished": true, "Ev Idx": 15481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740916857.905, "dur": 11.684,
    "args": {
      "Python parent id": 12, "Python id": 12582, "finished": true, "Ev Idx": 15482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916871.022, "dur": 15.167,
    "args": {
      "Python parent id": 12, "Python id": 12583, "finished": true, "Ev Idx": 15483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916871.342, "dur": 11.479,
    "args": {
      "Python parent id": 12583, "Python id": 12584, "finished": true, "Ev Idx": 15484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916871.491, "dur": 0.507,
    "args": {
      "Python parent id": 12584, "Python id": 12585, "finished": true, "Ev Idx": 15485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916871.689, "dur": 0.242,
    "args": {
      "Python parent id": 12585, "Python id": 12586, "finished": true, "Ev Idx": 15486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916871.785, "dur": 0.111,
    "args": {
      "Python parent id": 12586, "Python id": 12587, "finished": true, "Ev Idx": 15487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916872.269, "dur": 8.595,
    "args": {
      "Python parent id": 12584, "Python id": 12588, "finished": true, "Ev Idx": 15488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916872.355, "dur": 0.070,
    "args": {
      "Python parent id": 12588, "Python id": 12589, "finished": true, "Ev Idx": 15489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916872.533, "dur": 0.070,
    "args": {
      "Python parent id": 12588, "Python id": 12590, "finished": true, "Ev Idx": 15490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916872.817, "dur": 0.059,
    "args": {
      "Python parent id": 12588, "Python id": 12591, "finished": true, "Ev Idx": 15491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916873.337, "dur": 0.045,
    "args": {
      "Python parent id": 12588, "Python id": 12592, "finished": true, "Ev Idx": 15492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916873.639, "dur": 0.081,
    "args": {
      "Python parent id": 12588, "Python id": 12593, "finished": true, "Ev Idx": 15493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916874.052, "dur": 6.759,
    "args": {
      "Python parent id": 12588, "Python id": 12594, "finished": true, "Ev Idx": 15494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916874.148, "dur": 0.065,
    "args": {
      "Python parent id": 12594, "Python id": 12595, "finished": true, "Ev Idx": 15495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916874.380, "dur": 0.057,
    "args": {
      "Python parent id": 12594, "Python id": 12596, "finished": true, "Ev Idx": 15496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916874.619, "dur": 0.069,
    "args": {
      "Python parent id": 12594, "Python id": 12597, "finished": true, "Ev Idx": 15497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916875.009, "dur": 0.067,
    "args": {
      "Python parent id": 12594, "Python id": 12598, "finished": true, "Ev Idx": 15498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916875.357, "dur": 5.319,
    "args": {
      "Python parent id": 12594, "Python id": 12599, "finished": true, "Ev Idx": 15499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916875.606, "dur": 5.009,
    "args": {
      "Python parent id": 12599, "Python id": 12600, "finished": true, "Ev Idx": 15500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916875.788, "dur": 2.976,
    "args": {
      "Python parent id": 12600, "Python id": 12601, "finished": true, "Ev Idx": 15501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916875.995, "dur": 2.721,
    "args": {
      "Python parent id": 12601, "Python id": 12602, "finished": true, "Ev Idx": 15502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916876.124, "dur": 0.251,
    "args": {
      "Python parent id": 12602, "Python id": 12603, "finished": true, "Ev Idx": 15503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916876.244, "dur": 0.066,
    "args": {
      "Python parent id": 12603, "Python id": 12604, "finished": true, "Ev Idx": 15504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916876.521, "dur": 1.924,
    "args": {
      "Python parent id": 12602, "Python id": 12605, "finished": true, "Ev Idx": 15505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916876.687, "dur": 1.693,
    "args": {
      "Python parent id": 12605, "Python id": 12606, "finished": true, "Ev Idx": 15506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916876.873, "dur": 1.456,
    "args": {
      "Python parent id": 12606, "Python id": 12607, "finished": true, "Ev Idx": 15507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916877.012, "dur": 1.119,
    "args": {
      "Python parent id": 12607, "Python id": 12608, "finished": true, "Ev Idx": 15508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916877.207, "dur": 0.478,
    "args": {
      "Python parent id": 12608, "Python id": 12609, "finished": true, "Ev Idx": 15509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916877.290, "dur": 0.075,
    "args": {
      "Python parent id": 12609, "Python id": 12610, "finished": true, "Ev Idx": 15510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916877.517, "dur": 0.125,
    "args": {
      "Python parent id": 12609, "Python id": 12611, "finished": true, "Ev Idx": 15511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916878.606, "dur": 0.072,
    "args": {
      "Python parent id": 12602, "Python id": 12612, "finished": true, "Ev Idx": 15512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916878.928, "dur": 0.100,
    "args": {
      "Python parent id": 12600, "Python id": 12613, "finished": true, "Ev Idx": 15513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916879.368, "dur": 1.201,
    "args": {
      "Python parent id": 12600, "Python id": 12614, "finished": true, "Ev Idx": 15514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916879.516, "dur": 1.009,
    "args": {
      "Python parent id": 12614, "Python id": 12615, "finished": true, "Ev Idx": 15515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916879.658, "dur": 0.412,
    "args": {
      "Python parent id": 12615, "Python id": 12616, "finished": true, "Ev Idx": 15516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916879.833, "dur": 0.178,
    "args": {
      "Python parent id": 12616, "Python id": 12617, "finished": true, "Ev Idx": 15517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916879.917, "dur": 0.059,
    "args": {
      "Python parent id": 12617, "Python id": 12618, "finished": true, "Ev Idx": 15518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916880.199, "dur": 0.299,
    "args": {
      "Python parent id": 12615, "Python id": 12619, "finished": true, "Ev Idx": 15519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916880.986, "dur": 0.150,
    "args": {
      "Python parent id": 12584, "Python id": 12620, "finished": true, "Ev Idx": 15520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916881.753, "dur": 0.923,
    "args": {
      "Python parent id": 12584, "Python id": 12621, "finished": true, "Ev Idx": 15521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916882.204, "dur": 0.420,
    "args": {
      "Python parent id": 12621, "Python id": 12622, "finished": true, "Ev Idx": 15522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916883.098, "dur": 3.016,
    "args": {
      "Python parent id": 12583, "Python id": 12623, "finished": true, "Ev Idx": 15523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740916887.142, "dur": 96.038,
    "args": {
      "Python parent id": 12, "Python id": 12624, "finished": true, "Ev Idx": 15524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916887.556, "dur": 0.465,
    "args": {
      "Python parent id": 12624, "Python id": 12625, "finished": true, "Ev Idx": 15525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916887.846, "dur": 0.109,
    "args": {
      "Python parent id": 12625, "Python id": 12626, "finished": true, "Ev Idx": 15526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916888.306, "dur": 0.412,
    "args": {
      "Python parent id": 12624, "Python id": 12627, "finished": true, "Ev Idx": 15527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740916888.960, "dur": 0.296,
    "args": {
      "Python parent id": 12624, "Python id": 12628, "finished": true, "Ev Idx": 15528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740916889.504, "dur": 1.654,
    "args": {
      "Python parent id": 12624, "Python id": 12629, "finished": true, "Ev Idx": 15529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740916891.864, "dur": 91.211,
    "args": {
      "Python parent id": 12624, "Python id": 12630, "finished": true, "Ev Idx": 15530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740916892.738, "dur": 90.111,
    "args": {
      "Python parent id": 12630, "Python id": 12631, "finished": true, "Ev Idx": 15531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916893.057, "dur": 89.713,
    "args": {
      "Python parent id": 12631, "Python id": 12632, "finished": true, "Ev Idx": 15532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740916899.066, "dur": 82.354,
    "args": {
      "Python parent id": 12632, "Python id": 12633, "finished": true, "Ev Idx": 15533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740916899.408, "dur": 81.805,
    "args": {
      "Python parent id": 12633, "Python id": 12634, "finished": true, "Ev Idx": 15534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740916900.059, "dur": 81.077,
    "args": {
      "Python parent id": 12634, "Python id": 12635, "finished": true, "Ev Idx": 15535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916900.595, "dur": 0.423,
    "args": {
      "Python parent id": 12635, "Python id": 12636, "finished": true, "Ev Idx": 15536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916900.863, "dur": 0.077,
    "args": {
      "Python parent id": 12636, "Python id": 12637, "finished": true, "Ev Idx": 15537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740916901.382, "dur": 0.328,
    "args": {
      "Python parent id": 12635, "Python id": 12638, "finished": true, "Ev Idx": 15538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740916901.602, "dur": 0.059,
    "args": {
      "Python parent id": 12638, "Python id": 12639, "finished": true, "Ev Idx": 15539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916902.039, "dur": 78.982,
    "args": {
      "Python parent id": 12635, "Python id": 12640, "finished": true, "Ev Idx": 15540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740916902.853, "dur": 76.901,
    "args": {
      "Python parent id": 12640, "Python id": 12641, "finished": true, "Ev Idx": 15541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740916910.579, "dur": 20.358,
    "args": {
      "Python parent id": 12641, "Python id": 12642, "finished": true, "Ev Idx": 15542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740916910.721, "dur": 0.208,
    "args": {
      "Python parent id": 12642, "Python id": 12643, "finished": true, "Ev Idx": 15543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740916911.215, "dur": 19.571,
    "args": {
      "Python parent id": 12642, "Python id": 12644, "finished": true, "Ev Idx": 15544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740916911.350, "dur": 9.478,
    "args": {
      "Python parent id": 12644, "Python id": 12645, "finished": true, "Ev Idx": 15545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916938.995, "dur": 7.610,
    "args": {
      "Python parent id": 12641, "Python id": 12646, "finished": true, "Ev Idx": 15546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740916955.185, "dur": 8.613,
    "args": {
      "Python parent id": 12641, "Python id": 12647, "finished": true, "Ev Idx": 15547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740916984.525, "dur": 17.643,
    "args": {
      "Python parent id": 12, "Python id": 12648, "finished": true, "Ev Idx": 15548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740916984.923, "dur": 13.399,
    "args": {
      "Python parent id": 12648, "Python id": 12649, "finished": true, "Ev Idx": 15549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916985.148, "dur": 0.528,
    "args": {
      "Python parent id": 12649, "Python id": 12650, "finished": true, "Ev Idx": 15550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916985.316, "dur": 0.291,
    "args": {
      "Python parent id": 12650, "Python id": 12651, "finished": true, "Ev Idx": 15551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916985.459, "dur": 0.113,
    "args": {
      "Python parent id": 12651, "Python id": 12652, "finished": true, "Ev Idx": 15552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916986.008, "dur": 9.754,
    "args": {
      "Python parent id": 12649, "Python id": 12653, "finished": true, "Ev Idx": 15553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916986.095, "dur": 0.069,
    "args": {
      "Python parent id": 12653, "Python id": 12654, "finished": true, "Ev Idx": 15554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916986.281, "dur": 0.061,
    "args": {
      "Python parent id": 12653, "Python id": 12655, "finished": true, "Ev Idx": 15555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916986.501, "dur": 0.063,
    "args": {
      "Python parent id": 12653, "Python id": 12656, "finished": true, "Ev Idx": 15556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916986.977, "dur": 0.083,
    "args": {
      "Python parent id": 12653, "Python id": 12657, "finished": true, "Ev Idx": 15557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916987.353, "dur": 0.079,
    "args": {
      "Python parent id": 12653, "Python id": 12658, "finished": true, "Ev Idx": 15558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916987.800, "dur": 7.909,
    "args": {
      "Python parent id": 12653, "Python id": 12659, "finished": true, "Ev Idx": 15559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916987.939, "dur": 0.064,
    "args": {
      "Python parent id": 12659, "Python id": 12660, "finished": true, "Ev Idx": 15560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916988.190, "dur": 0.061,
    "args": {
      "Python parent id": 12659, "Python id": 12661, "finished": true, "Ev Idx": 15561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916988.463, "dur": 0.070,
    "args": {
      "Python parent id": 12659, "Python id": 12662, "finished": true, "Ev Idx": 15562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740916988.910, "dur": 0.067,
    "args": {
      "Python parent id": 12659, "Python id": 12663, "finished": true, "Ev Idx": 15563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740916989.346, "dur": 6.262,
    "args": {
      "Python parent id": 12659, "Python id": 12664, "finished": true, "Ev Idx": 15564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740916989.618, "dur": 5.913,
    "args": {
      "Python parent id": 12664, "Python id": 12665, "finished": true, "Ev Idx": 15565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740916989.763, "dur": 3.904,
    "args": {
      "Python parent id": 12665, "Python id": 12666, "finished": true, "Ev Idx": 15566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740916990.022, "dur": 3.574,
    "args": {
      "Python parent id": 12666, "Python id": 12667, "finished": true, "Ev Idx": 15567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740916990.159, "dur": 0.256,
    "args": {
      "Python parent id": 12667, "Python id": 12668, "finished": true, "Ev Idx": 15568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740916990.287, "dur": 0.067,
    "args": {
      "Python parent id": 12668, "Python id": 12669, "finished": true, "Ev Idx": 15569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740916990.587, "dur": 2.584,
    "args": {
      "Python parent id": 12667, "Python id": 12670, "finished": true, "Ev Idx": 15570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740916990.864, "dur": 2.200,
    "args": {
      "Python parent id": 12670, "Python id": 12671, "finished": true, "Ev Idx": 15571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740916991.147, "dur": 1.835,
    "args": {
      "Python parent id": 12671, "Python id": 12672, "finished": true, "Ev Idx": 15572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740916991.320, "dur": 1.362,
    "args": {
      "Python parent id": 12672, "Python id": 12673, "finished": true, "Ev Idx": 15573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740916991.505, "dur": 0.623,
    "args": {
      "Python parent id": 12673, "Python id": 12674, "finished": true, "Ev Idx": 15574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740916991.616, "dur": 0.068,
    "args": {
      "Python parent id": 12674, "Python id": 12675, "finished": true, "Ev Idx": 15575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740916991.891, "dur": 0.165,
    "args": {
      "Python parent id": 12674, "Python id": 12676, "finished": true, "Ev Idx": 15576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740916993.366, "dur": 0.184,
    "args": {
      "Python parent id": 12667, "Python id": 12677, "finished": true, "Ev Idx": 15577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740916993.856, "dur": 0.088,
    "args": {
      "Python parent id": 12665, "Python id": 12678, "finished": true, "Ev Idx": 15578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740916994.248, "dur": 1.240,
    "args": {
      "Python parent id": 12665, "Python id": 12679, "finished": true, "Ev Idx": 15579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740916994.404, "dur": 1.030,
    "args": {
      "Python parent id": 12679, "Python id": 12680, "finished": true, "Ev Idx": 15580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740916994.566, "dur": 0.408,
    "args": {
      "Python parent id": 12680, "Python id": 12681, "finished": true, "Ev Idx": 15581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740916994.708, "dur": 0.187,
    "args": {
      "Python parent id": 12681, "Python id": 12682, "finished": true, "Ev Idx": 15582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740916994.803, "dur": 0.059,
    "args": {
      "Python parent id": 12682, "Python id": 12683, "finished": true, "Ev Idx": 15583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740916995.092, "dur": 0.272,
    "args": {
      "Python parent id": 12680, "Python id": 12684, "finished": true, "Ev Idx": 15584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740916995.912, "dur": 0.303,
    "args": {
      "Python parent id": 12649, "Python id": 12685, "finished": true, "Ev Idx": 15585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740916996.960, "dur": 1.219,
    "args": {
      "Python parent id": 12649, "Python id": 12686, "finished": true, "Ev Idx": 15586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740916997.428, "dur": 0.699,
    "args": {
      "Python parent id": 12686, "Python id": 12687, "finished": true, "Ev Idx": 15587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740916998.680, "dur": 3.376,
    "args": {
      "Python parent id": 12648, "Python id": 12688, "finished": true, "Ev Idx": 15588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917002.733, "dur": 8.162,
    "args": {
      "Python parent id": 12, "Python id": 12689, "finished": true, "Ev Idx": 15589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740917011.415, "dur": 0.079,
    "args": {
      "Python parent id": 12, "Python id": 12690, "finished": true, "Ev Idx": 15590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917012.140, "dur": 0.215,
    "args": {
      "Python parent id": 12, "Python id": 12691, "finished": true, "Ev Idx": 15591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917012.895, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 12692, "finished": true, "Ev Idx": 15592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740917013.353, "dur": 10.214,
    "args": {
      "Python parent id": 12, "Python id": 12693, "finished": true, "Ev Idx": 15593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917025.010, "dur": 15.745,
    "args": {
      "Python parent id": 12, "Python id": 12694, "finished": true, "Ev Idx": 15594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917025.360, "dur": 11.946,
    "args": {
      "Python parent id": 12694, "Python id": 12695, "finished": true, "Ev Idx": 15595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917025.599, "dur": 0.556,
    "args": {
      "Python parent id": 12695, "Python id": 12696, "finished": true, "Ev Idx": 15596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917025.775, "dur": 0.287,
    "args": {
      "Python parent id": 12696, "Python id": 12697, "finished": true, "Ev Idx": 15597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917025.919, "dur": 0.104,
    "args": {
      "Python parent id": 12697, "Python id": 12698, "finished": true, "Ev Idx": 15598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917026.413, "dur": 8.894,
    "args": {
      "Python parent id": 12695, "Python id": 12699, "finished": true, "Ev Idx": 15599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917026.502, "dur": 0.075,
    "args": {
      "Python parent id": 12699, "Python id": 12700, "finished": true, "Ev Idx": 15600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917026.714, "dur": 0.059,
    "args": {
      "Python parent id": 12699, "Python id": 12701, "finished": true, "Ev Idx": 15601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917026.933, "dur": 0.063,
    "args": {
      "Python parent id": 12699, "Python id": 12702, "finished": true, "Ev Idx": 15602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917027.456, "dur": 0.033,
    "args": {
      "Python parent id": 12699, "Python id": 12703, "finished": true, "Ev Idx": 15603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917027.775, "dur": 0.079,
    "args": {
      "Python parent id": 12699, "Python id": 12704, "finished": true, "Ev Idx": 15604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917028.235, "dur": 6.981,
    "args": {
      "Python parent id": 12699, "Python id": 12705, "finished": true, "Ev Idx": 15605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917028.331, "dur": 0.065,
    "args": {
      "Python parent id": 12705, "Python id": 12706, "finished": true, "Ev Idx": 15606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917028.548, "dur": 0.064,
    "args": {
      "Python parent id": 12705, "Python id": 12707, "finished": true, "Ev Idx": 15607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917028.838, "dur": 0.070,
    "args": {
      "Python parent id": 12705, "Python id": 12708, "finished": true, "Ev Idx": 15608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917029.304, "dur": 0.036,
    "args": {
      "Python parent id": 12705, "Python id": 12709, "finished": true, "Ev Idx": 15609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917029.664, "dur": 5.446,
    "args": {
      "Python parent id": 12705, "Python id": 12710, "finished": true, "Ev Idx": 15610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917029.891, "dur": 5.143,
    "args": {
      "Python parent id": 12710, "Python id": 12711, "finished": true, "Ev Idx": 15611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917030.028, "dur": 3.229,
    "args": {
      "Python parent id": 12711, "Python id": 12712, "finished": true, "Ev Idx": 15612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917030.247, "dur": 2.962,
    "args": {
      "Python parent id": 12712, "Python id": 12713, "finished": true, "Ev Idx": 15613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917030.381, "dur": 0.286,
    "args": {
      "Python parent id": 12713, "Python id": 12714, "finished": true, "Ev Idx": 15614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917030.519, "dur": 0.068,
    "args": {
      "Python parent id": 12714, "Python id": 12715, "finished": true, "Ev Idx": 15615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917030.813, "dur": 2.055,
    "args": {
      "Python parent id": 12713, "Python id": 12716, "finished": true, "Ev Idx": 15616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.007, "dur": 1.800,
    "args": {
      "Python parent id": 12716, "Python id": 12717, "finished": true, "Ev Idx": 15617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.178, "dur": 1.580,
    "args": {
      "Python parent id": 12717, "Python id": 12718, "finished": true, "Ev Idx": 15618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.333, "dur": 1.175,
    "args": {
      "Python parent id": 12718, "Python id": 12719, "finished": true, "Ev Idx": 15619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.541, "dur": 0.538,
    "args": {
      "Python parent id": 12719, "Python id": 12720, "finished": true, "Ev Idx": 15620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.627, "dur": 0.077,
    "args": {
      "Python parent id": 12720, "Python id": 12721, "finished": true, "Ev Idx": 15621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917031.867, "dur": 0.147,
    "args": {
      "Python parent id": 12720, "Python id": 12722, "finished": true, "Ev Idx": 15622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917033.057, "dur": 0.113,
    "args": {
      "Python parent id": 12713, "Python id": 12723, "finished": true, "Ev Idx": 15623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917033.414, "dur": 0.090,
    "args": {
      "Python parent id": 12711, "Python id": 12724, "finished": true, "Ev Idx": 15624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917033.794, "dur": 1.194,
    "args": {
      "Python parent id": 12711, "Python id": 12725, "finished": true, "Ev Idx": 15625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917033.940, "dur": 1.004,
    "args": {
      "Python parent id": 12725, "Python id": 12726, "finished": true, "Ev Idx": 15626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917034.085, "dur": 0.412,
    "args": {
      "Python parent id": 12726, "Python id": 12727, "finished": true, "Ev Idx": 15627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917034.248, "dur": 0.191,
    "args": {
      "Python parent id": 12727, "Python id": 12728, "finished": true, "Ev Idx": 15628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917034.342, "dur": 0.065,
    "args": {
      "Python parent id": 12728, "Python id": 12729, "finished": true, "Ev Idx": 15629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917034.629, "dur": 0.275,
    "args": {
      "Python parent id": 12726, "Python id": 12730, "finished": true, "Ev Idx": 15630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917035.449, "dur": 0.147,
    "args": {
      "Python parent id": 12695, "Python id": 12731, "finished": true, "Ev Idx": 15631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917036.225, "dur": 0.940,
    "args": {
      "Python parent id": 12695, "Python id": 12732, "finished": true, "Ev Idx": 15632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917036.631, "dur": 0.483,
    "args": {
      "Python parent id": 12732, "Python id": 12733, "finished": true, "Ev Idx": 15633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917037.614, "dur": 3.086,
    "args": {
      "Python parent id": 12694, "Python id": 12734, "finished": true, "Ev Idx": 15634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740917041.696, "dur": 136.133,
    "args": {
      "Python parent id": 12, "Python id": 12735, "finished": true, "Ev Idx": 15635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917042.024, "dur": 0.525,
    "args": {
      "Python parent id": 12735, "Python id": 12736, "finished": true, "Ev Idx": 15636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917042.345, "dur": 0.107,
    "args": {
      "Python parent id": 12736, "Python id": 12737, "finished": true, "Ev Idx": 15637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917042.871, "dur": 0.408,
    "args": {
      "Python parent id": 12735, "Python id": 12738, "finished": true, "Ev Idx": 15638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917043.494, "dur": 0.269,
    "args": {
      "Python parent id": 12735, "Python id": 12739, "finished": true, "Ev Idx": 15639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740917043.982, "dur": 1.564,
    "args": {
      "Python parent id": 12735, "Python id": 12740, "finished": true, "Ev Idx": 15640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740917046.160, "dur": 131.520,
    "args": {
      "Python parent id": 12735, "Python id": 12741, "finished": true, "Ev Idx": 15641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740917047.047, "dur": 130.415,
    "args": {
      "Python parent id": 12741, "Python id": 12742, "finished": true, "Ev Idx": 15642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917047.354, "dur": 130.006,
    "args": {
      "Python parent id": 12742, "Python id": 12743, "finished": true, "Ev Idx": 15643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917059.904, "dur": 115.721,
    "args": {
      "Python parent id": 12743, "Python id": 12744, "finished": true, "Ev Idx": 15644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740917060.350, "dur": 115.144,
    "args": {
      "Python parent id": 12744, "Python id": 12745, "finished": true, "Ev Idx": 15645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740917061.094, "dur": 114.276,
    "args": {
      "Python parent id": 12745, "Python id": 12746, "finished": true, "Ev Idx": 15646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917061.805, "dur": 0.544,
    "args": {
      "Python parent id": 12746, "Python id": 12747, "finished": true, "Ev Idx": 15647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917062.145, "dur": 0.100,
    "args": {
      "Python parent id": 12747, "Python id": 12748, "finished": true, "Ev Idx": 15648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917062.662, "dur": 0.349,
    "args": {
      "Python parent id": 12746, "Python id": 12749, "finished": true, "Ev Idx": 15649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917062.911, "dur": 0.054,
    "args": {
      "Python parent id": 12749, "Python id": 12750, "finished": true, "Ev Idx": 15650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917063.381, "dur": 111.877,
    "args": {
      "Python parent id": 12746, "Python id": 12751, "finished": true, "Ev Idx": 15651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917064.276, "dur": 109.381,
    "args": {
      "Python parent id": 12751, "Python id": 12752, "finished": true, "Ev Idx": 15652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740917077.932, "dur": 26.477,
    "args": {
      "Python parent id": 12752, "Python id": 12753, "finished": true, "Ev Idx": 15653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740917078.134, "dur": 0.152,
    "args": {
      "Python parent id": 12753, "Python id": 12754, "finished": true, "Ev Idx": 15654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740917078.669, "dur": 25.515,
    "args": {
      "Python parent id": 12753, "Python id": 12755, "finished": true, "Ev Idx": 15655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917078.823, "dur": 10.941,
    "args": {
      "Python parent id": 12755, "Python id": 12756, "finished": true, "Ev Idx": 15656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917114.078, "dur": 8.502,
    "args": {
      "Python parent id": 12752, "Python id": 12757, "finished": true, "Ev Idx": 15657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917146.603, "dur": 10.150,
    "args": {
      "Python parent id": 12752, "Python id": 12758, "finished": true, "Ev Idx": 15658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917179.686, "dur": 21.591,
    "args": {
      "Python parent id": 12, "Python id": 12759, "finished": true, "Ev Idx": 15659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917180.180, "dur": 15.936,
    "args": {
      "Python parent id": 12759, "Python id": 12760, "finished": true, "Ev Idx": 15660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917180.416, "dur": 0.649,
    "args": {
      "Python parent id": 12760, "Python id": 12761, "finished": true, "Ev Idx": 15661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917180.572, "dur": 0.404,
    "args": {
      "Python parent id": 12761, "Python id": 12762, "finished": true, "Ev Idx": 15662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917180.757, "dur": 0.171,
    "args": {
      "Python parent id": 12762, "Python id": 12763, "finished": true, "Ev Idx": 15663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917181.445, "dur": 11.447,
    "args": {
      "Python parent id": 12760, "Python id": 12764, "finished": true, "Ev Idx": 15664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917181.625, "dur": 0.090,
    "args": {
      "Python parent id": 12764, "Python id": 12765, "finished": true, "Ev Idx": 15665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917181.830, "dur": 0.081,
    "args": {
      "Python parent id": 12764, "Python id": 12766, "finished": true, "Ev Idx": 15666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917182.058, "dur": 0.057,
    "args": {
      "Python parent id": 12764, "Python id": 12767, "finished": true, "Ev Idx": 15667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917182.738, "dur": 0.046,
    "args": {
      "Python parent id": 12764, "Python id": 12768, "finished": true, "Ev Idx": 15668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917183.124, "dur": 0.097,
    "args": {
      "Python parent id": 12764, "Python id": 12769, "finished": true, "Ev Idx": 15669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917183.616, "dur": 9.221,
    "args": {
      "Python parent id": 12764, "Python id": 12770, "finished": true, "Ev Idx": 15670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917183.710, "dur": 0.063,
    "args": {
      "Python parent id": 12770, "Python id": 12771, "finished": true, "Ev Idx": 15671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917183.982, "dur": 0.058,
    "args": {
      "Python parent id": 12770, "Python id": 12772, "finished": true, "Ev Idx": 15672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917184.238, "dur": 0.064,
    "args": {
      "Python parent id": 12770, "Python id": 12773, "finished": true, "Ev Idx": 15673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917184.697, "dur": 0.036,
    "args": {
      "Python parent id": 12770, "Python id": 12774, "finished": true, "Ev Idx": 15674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917185.045, "dur": 7.644,
    "args": {
      "Python parent id": 12770, "Python id": 12775, "finished": true, "Ev Idx": 15675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917185.349, "dur": 7.263,
    "args": {
      "Python parent id": 12775, "Python id": 12776, "finished": true, "Ev Idx": 15676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917185.524, "dur": 4.750,
    "args": {
      "Python parent id": 12776, "Python id": 12777, "finished": true, "Ev Idx": 15677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917185.741, "dur": 4.445,
    "args": {
      "Python parent id": 12777, "Python id": 12778, "finished": true, "Ev Idx": 15678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917185.869, "dur": 0.365,
    "args": {
      "Python parent id": 12778, "Python id": 12779, "finished": true, "Ev Idx": 15679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917186.036, "dur": 0.089,
    "args": {
      "Python parent id": 12779, "Python id": 12780, "finished": true, "Ev Idx": 15680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917186.429, "dur": 3.220,
    "args": {
      "Python parent id": 12778, "Python id": 12781, "finished": true, "Ev Idx": 15681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917186.677, "dur": 2.808,
    "args": {
      "Python parent id": 12781, "Python id": 12782, "finished": true, "Ev Idx": 15682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917186.944, "dur": 2.453,
    "args": {
      "Python parent id": 12782, "Python id": 12783, "finished": true, "Ev Idx": 15683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917187.128, "dur": 1.855,
    "args": {
      "Python parent id": 12783, "Python id": 12784, "finished": true, "Ev Idx": 15684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917187.379, "dur": 0.789,
    "args": {
      "Python parent id": 12784, "Python id": 12785, "finished": true, "Ev Idx": 15685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917187.486, "dur": 0.067,
    "args": {
      "Python parent id": 12785, "Python id": 12786, "finished": true, "Ev Idx": 15686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917187.770, "dur": 0.308,
    "args": {
      "Python parent id": 12785, "Python id": 12787, "finished": true, "Ev Idx": 15687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917189.869, "dur": 0.236,
    "args": {
      "Python parent id": 12778, "Python id": 12788, "finished": true, "Ev Idx": 15688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917190.525, "dur": 0.151,
    "args": {
      "Python parent id": 12776, "Python id": 12789, "finished": true, "Ev Idx": 15689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917191.042, "dur": 1.507,
    "args": {
      "Python parent id": 12776, "Python id": 12790, "finished": true, "Ev Idx": 15690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917191.258, "dur": 1.238,
    "args": {
      "Python parent id": 12790, "Python id": 12791, "finished": true, "Ev Idx": 15691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917191.422, "dur": 0.483,
    "args": {
      "Python parent id": 12791, "Python id": 12792, "finished": true, "Ev Idx": 15692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917191.572, "dur": 0.250,
    "args": {
      "Python parent id": 12792, "Python id": 12793, "finished": true, "Ev Idx": 15693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917191.687, "dur": 0.099,
    "args": {
      "Python parent id": 12793, "Python id": 12794, "finished": true, "Ev Idx": 15694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917192.068, "dur": 0.362,
    "args": {
      "Python parent id": 12791, "Python id": 12795, "finished": true, "Ev Idx": 15695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917193.058, "dur": 0.338,
    "args": {
      "Python parent id": 12760, "Python id": 12796, "finished": true, "Ev Idx": 15696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917194.282, "dur": 1.618,
    "args": {
      "Python parent id": 12760, "Python id": 12797, "finished": true, "Ev Idx": 15697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917195.004, "dur": 0.802,
    "args": {
      "Python parent id": 12797, "Python id": 12798, "finished": true, "Ev Idx": 15698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917196.549, "dur": 4.597,
    "args": {
      "Python parent id": 12759, "Python id": 12799, "finished": true, "Ev Idx": 15699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917201.942, "dur": 10.500,
    "args": {
      "Python parent id": 12, "Python id": 12800, "finished": true, "Ev Idx": 15700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740917213.010, "dur": 0.137,
    "args": {
      "Python parent id": 12, "Python id": 12801, "finished": true, "Ev Idx": 15701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917213.924, "dur": 0.261,
    "args": {
      "Python parent id": 12, "Python id": 12802, "finished": true, "Ev Idx": 15702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917214.730, "dur": 0.071,
    "args": {
      "Python parent id": 12, "Python id": 12803, "finished": true, "Ev Idx": 15703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740917215.234, "dur": 11.711,
    "args": {
      "Python parent id": 12, "Python id": 12804, "finished": true, "Ev Idx": 15704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917228.285, "dur": 17.987,
    "args": {
      "Python parent id": 12, "Python id": 12805, "finished": true, "Ev Idx": 15705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917228.667, "dur": 13.126,
    "args": {
      "Python parent id": 12805, "Python id": 12806, "finished": true, "Ev Idx": 15706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917228.828, "dur": 0.487,
    "args": {
      "Python parent id": 12806, "Python id": 12807, "finished": true, "Ev Idx": 15707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917228.986, "dur": 0.264,
    "args": {
      "Python parent id": 12807, "Python id": 12808, "finished": true, "Ev Idx": 15708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917229.086, "dur": 0.127,
    "args": {
      "Python parent id": 12808, "Python id": 12809, "finished": true, "Ev Idx": 15709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917229.655, "dur": 9.880,
    "args": {
      "Python parent id": 12806, "Python id": 12810, "finished": true, "Ev Idx": 15710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917229.762, "dur": 0.071,
    "args": {
      "Python parent id": 12810, "Python id": 12811, "finished": true, "Ev Idx": 15711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917229.943, "dur": 0.067,
    "args": {
      "Python parent id": 12810, "Python id": 12812, "finished": true, "Ev Idx": 15712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917230.157, "dur": 0.064,
    "args": {
      "Python parent id": 12810, "Python id": 12813, "finished": true, "Ev Idx": 15713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917230.605, "dur": 0.072,
    "args": {
      "Python parent id": 12810, "Python id": 12814, "finished": true, "Ev Idx": 15714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917230.974, "dur": 0.079,
    "args": {
      "Python parent id": 12810, "Python id": 12815, "finished": true, "Ev Idx": 15715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917231.415, "dur": 8.046,
    "args": {
      "Python parent id": 12810, "Python id": 12816, "finished": true, "Ev Idx": 15716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917231.539, "dur": 0.065,
    "args": {
      "Python parent id": 12816, "Python id": 12817, "finished": true, "Ev Idx": 15717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917231.780, "dur": 0.566,
    "args": {
      "Python parent id": 12816, "Python id": 12818, "finished": true, "Ev Idx": 15718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917232.573, "dur": 0.057,
    "args": {
      "Python parent id": 12816, "Python id": 12819, "finished": true, "Ev Idx": 15719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917233.034, "dur": 0.036,
    "args": {
      "Python parent id": 12816, "Python id": 12820, "finished": true, "Ev Idx": 15720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917233.405, "dur": 5.953,
    "args": {
      "Python parent id": 12816, "Python id": 12821, "finished": true, "Ev Idx": 15721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917233.673, "dur": 5.579,
    "args": {
      "Python parent id": 12821, "Python id": 12822, "finished": true, "Ev Idx": 15722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917233.799, "dur": 3.631,
    "args": {
      "Python parent id": 12822, "Python id": 12823, "finished": true, "Ev Idx": 15723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917234.058, "dur": 3.278,
    "args": {
      "Python parent id": 12823, "Python id": 12824, "finished": true, "Ev Idx": 15724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917234.191, "dur": 0.334,
    "args": {
      "Python parent id": 12824, "Python id": 12825, "finished": true, "Ev Idx": 15725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917234.347, "dur": 0.094,
    "args": {
      "Python parent id": 12825, "Python id": 12826, "finished": true, "Ev Idx": 15726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917234.759, "dur": 2.294,
    "args": {
      "Python parent id": 12824, "Python id": 12827, "finished": true, "Ev Idx": 15727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917234.952, "dur": 1.990,
    "args": {
      "Python parent id": 12827, "Python id": 12828, "finished": true, "Ev Idx": 15728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917235.194, "dur": 1.681,
    "args": {
      "Python parent id": 12828, "Python id": 12829, "finished": true, "Ev Idx": 15729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917235.354, "dur": 1.303,
    "args": {
      "Python parent id": 12829, "Python id": 12830, "finished": true, "Ev Idx": 15730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917235.575, "dur": 0.527,
    "args": {
      "Python parent id": 12830, "Python id": 12831, "finished": true, "Ev Idx": 15731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917235.681, "dur": 0.043,
    "args": {
      "Python parent id": 12831, "Python id": 12832, "finished": true, "Ev Idx": 15732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917235.883, "dur": 0.159,
    "args": {
      "Python parent id": 12831, "Python id": 12833, "finished": true, "Ev Idx": 15733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917237.222, "dur": 0.076,
    "args": {
      "Python parent id": 12824, "Python id": 12834, "finished": true, "Ev Idx": 15734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917237.616, "dur": 0.090,
    "args": {
      "Python parent id": 12822, "Python id": 12835, "finished": true, "Ev Idx": 15735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.005, "dur": 1.188,
    "args": {
      "Python parent id": 12822, "Python id": 12836, "finished": true, "Ev Idx": 15736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.173, "dur": 0.974,
    "args": {
      "Python parent id": 12836, "Python id": 12837, "finished": true, "Ev Idx": 15737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.310, "dur": 0.403,
    "args": {
      "Python parent id": 12837, "Python id": 12838, "finished": true, "Ev Idx": 15738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.452, "dur": 0.202,
    "args": {
      "Python parent id": 12838, "Python id": 12839, "finished": true, "Ev Idx": 15739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.559, "dur": 0.058,
    "args": {
      "Python parent id": 12839, "Python id": 12840, "finished": true, "Ev Idx": 15740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917238.845, "dur": 0.261,
    "args": {
      "Python parent id": 12837, "Python id": 12841, "finished": true, "Ev Idx": 15741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917239.692, "dur": 0.182,
    "args": {
      "Python parent id": 12806, "Python id": 12842, "finished": true, "Ev Idx": 15742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917240.591, "dur": 1.061,
    "args": {
      "Python parent id": 12806, "Python id": 12843, "finished": true, "Ev Idx": 15743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917241.051, "dur": 0.506,
    "args": {
      "Python parent id": 12843, "Python id": 12844, "finished": true, "Ev Idx": 15744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917242.142, "dur": 4.039,
    "args": {
      "Python parent id": 12805, "Python id": 12845, "finished": true, "Ev Idx": 15745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740917247.270, "dur": 99.448,
    "args": {
      "Python parent id": 12, "Python id": 12846, "finished": true, "Ev Idx": 15746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917247.689, "dur": 0.572,
    "args": {
      "Python parent id": 12846, "Python id": 12847, "finished": true, "Ev Idx": 15747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917248.062, "dur": 0.116,
    "args": {
      "Python parent id": 12847, "Python id": 12848, "finished": true, "Ev Idx": 15748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917248.564, "dur": 0.537,
    "args": {
      "Python parent id": 12846, "Python id": 12849, "finished": true, "Ev Idx": 15749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917249.377, "dur": 0.319,
    "args": {
      "Python parent id": 12846, "Python id": 12850, "finished": true, "Ev Idx": 15750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740917249.933, "dur": 1.815,
    "args": {
      "Python parent id": 12846, "Python id": 12851, "finished": true, "Ev Idx": 15751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740917252.423, "dur": 94.155,
    "args": {
      "Python parent id": 12846, "Python id": 12852, "finished": true, "Ev Idx": 15752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740917253.629, "dur": 92.728,
    "args": {
      "Python parent id": 12852, "Python id": 12853, "finished": true, "Ev Idx": 15753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917254.030, "dur": 92.239,
    "args": {
      "Python parent id": 12853, "Python id": 12854, "finished": true, "Ev Idx": 15754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917260.366, "dur": 84.384,
    "args": {
      "Python parent id": 12854, "Python id": 12855, "finished": true, "Ev Idx": 15755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740917260.658, "dur": 84.007,
    "args": {
      "Python parent id": 12855, "Python id": 12856, "finished": true, "Ev Idx": 15756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740917261.322, "dur": 83.130,
    "args": {
      "Python parent id": 12856, "Python id": 12857, "finished": true, "Ev Idx": 15757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917261.905, "dur": 0.460,
    "args": {
      "Python parent id": 12857, "Python id": 12858, "finished": true, "Ev Idx": 15758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917262.205, "dur": 0.105,
    "args": {
      "Python parent id": 12858, "Python id": 12859, "finished": true, "Ev Idx": 15759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917262.701, "dur": 0.370,
    "args": {
      "Python parent id": 12857, "Python id": 12860, "finished": true, "Ev Idx": 15760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917262.973, "dur": 0.046,
    "args": {
      "Python parent id": 12860, "Python id": 12861, "finished": true, "Ev Idx": 15761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917263.453, "dur": 80.905,
    "args": {
      "Python parent id": 12857, "Python id": 12862, "finished": true, "Ev Idx": 15762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917264.178, "dur": 78.855,
    "args": {
      "Python parent id": 12862, "Python id": 12863, "finished": true, "Ev Idx": 15763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740917271.826, "dur": 19.871,
    "args": {
      "Python parent id": 12863, "Python id": 12864, "finished": true, "Ev Idx": 15764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740917272.004, "dur": 0.169,
    "args": {
      "Python parent id": 12864, "Python id": 12865, "finished": true, "Ev Idx": 15765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740917272.496, "dur": 19.036,
    "args": {
      "Python parent id": 12864, "Python id": 12866, "finished": true, "Ev Idx": 15766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917272.656, "dur": 7.691,
    "args": {
      "Python parent id": 12866, "Python id": 12867, "finished": true, "Ev Idx": 15767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917300.002, "dur": 7.318,
    "args": {
      "Python parent id": 12863, "Python id": 12868, "finished": true, "Ev Idx": 15768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917316.407, "dur": 8.843,
    "args": {
      "Python parent id": 12863, "Python id": 12869, "finished": true, "Ev Idx": 15769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917348.170, "dur": 18.841,
    "args": {
      "Python parent id": 12, "Python id": 12870, "finished": true, "Ev Idx": 15770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917348.548, "dur": 13.884,
    "args": {
      "Python parent id": 12870, "Python id": 12871, "finished": true, "Ev Idx": 15771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917348.759, "dur": 0.510,
    "args": {
      "Python parent id": 12871, "Python id": 12872, "finished": true, "Ev Idx": 15772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917348.916, "dur": 0.291,
    "args": {
      "Python parent id": 12872, "Python id": 12873, "finished": true, "Ev Idx": 15773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917349.040, "dur": 0.109,
    "args": {
      "Python parent id": 12873, "Python id": 12874, "finished": true, "Ev Idx": 15774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917349.684, "dur": 10.234,
    "args": {
      "Python parent id": 12871, "Python id": 12875, "finished": true, "Ev Idx": 15775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917349.839, "dur": 0.104,
    "args": {
      "Python parent id": 12875, "Python id": 12876, "finished": true, "Ev Idx": 15776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917350.058, "dur": 0.061,
    "args": {
      "Python parent id": 12875, "Python id": 12877, "finished": true, "Ev Idx": 15777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917350.272, "dur": 0.060,
    "args": {
      "Python parent id": 12875, "Python id": 12878, "finished": true, "Ev Idx": 15778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917350.776, "dur": 0.036,
    "args": {
      "Python parent id": 12875, "Python id": 12879, "finished": true, "Ev Idx": 15779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917351.094, "dur": 0.070,
    "args": {
      "Python parent id": 12875, "Python id": 12880, "finished": true, "Ev Idx": 15780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917351.567, "dur": 8.292,
    "args": {
      "Python parent id": 12875, "Python id": 12881, "finished": true, "Ev Idx": 15781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917351.658, "dur": 0.068,
    "args": {
      "Python parent id": 12881, "Python id": 12882, "finished": true, "Ev Idx": 15782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917351.899, "dur": 0.063,
    "args": {
      "Python parent id": 12881, "Python id": 12883, "finished": true, "Ev Idx": 15783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917352.148, "dur": 0.074,
    "args": {
      "Python parent id": 12881, "Python id": 12884, "finished": true, "Ev Idx": 15784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917352.627, "dur": 0.039,
    "args": {
      "Python parent id": 12881, "Python id": 12885, "finished": true, "Ev Idx": 15785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917352.971, "dur": 6.778,
    "args": {
      "Python parent id": 12881, "Python id": 12886, "finished": true, "Ev Idx": 15786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917353.287, "dur": 6.353,
    "args": {
      "Python parent id": 12886, "Python id": 12887, "finished": true, "Ev Idx": 15787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917353.464, "dur": 4.132,
    "args": {
      "Python parent id": 12887, "Python id": 12888, "finished": true, "Ev Idx": 15788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917353.707, "dur": 3.788,
    "args": {
      "Python parent id": 12888, "Python id": 12889, "finished": true, "Ev Idx": 15789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917353.927, "dur": 0.310,
    "args": {
      "Python parent id": 12889, "Python id": 12890, "finished": true, "Ev Idx": 15790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917354.053, "dur": 0.087,
    "args": {
      "Python parent id": 12890, "Python id": 12891, "finished": true, "Ev Idx": 15791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917354.434, "dur": 2.622,
    "args": {
      "Python parent id": 12889, "Python id": 12892, "finished": true, "Ev Idx": 15792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917354.672, "dur": 2.216,
    "args": {
      "Python parent id": 12892, "Python id": 12893, "finished": true, "Ev Idx": 15793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917354.880, "dur": 1.927,
    "args": {
      "Python parent id": 12893, "Python id": 12894, "finished": true, "Ev Idx": 15794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917355.102, "dur": 1.443,
    "args": {
      "Python parent id": 12894, "Python id": 12895, "finished": true, "Ev Idx": 15795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917355.321, "dur": 0.585,
    "args": {
      "Python parent id": 12895, "Python id": 12896, "finished": true, "Ev Idx": 15796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917355.411, "dur": 0.073,
    "args": {
      "Python parent id": 12896, "Python id": 12897, "finished": true, "Ev Idx": 15797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917355.664, "dur": 0.168,
    "args": {
      "Python parent id": 12896, "Python id": 12898, "finished": true, "Ev Idx": 15798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917357.235, "dur": 0.192,
    "args": {
      "Python parent id": 12889, "Python id": 12899, "finished": true, "Ev Idx": 15799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917357.790, "dur": 0.152,
    "args": {
      "Python parent id": 12887, "Python id": 12900, "finished": true, "Ev Idx": 15800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917358.278, "dur": 1.299,
    "args": {
      "Python parent id": 12887, "Python id": 12901, "finished": true, "Ev Idx": 15801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917358.472, "dur": 1.064,
    "args": {
      "Python parent id": 12901, "Python id": 12902, "finished": true, "Ev Idx": 15802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917358.610, "dur": 0.411,
    "args": {
      "Python parent id": 12902, "Python id": 12903, "finished": true, "Ev Idx": 15803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917358.746, "dur": 0.197,
    "args": {
      "Python parent id": 12903, "Python id": 12904, "finished": true, "Ev Idx": 15804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917358.851, "dur": 0.062,
    "args": {
      "Python parent id": 12904, "Python id": 12905, "finished": true, "Ev Idx": 15805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917359.176, "dur": 0.293,
    "args": {
      "Python parent id": 12902, "Python id": 12906, "finished": true, "Ev Idx": 15806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917360.080, "dur": 0.319,
    "args": {
      "Python parent id": 12871, "Python id": 12907, "finished": true, "Ev Idx": 15807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917361.045, "dur": 1.219,
    "args": {
      "Python parent id": 12871, "Python id": 12908, "finished": true, "Ev Idx": 15808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917361.587, "dur": 0.609,
    "args": {
      "Python parent id": 12908, "Python id": 12909, "finished": true, "Ev Idx": 15809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917362.773, "dur": 4.154,
    "args": {
      "Python parent id": 12870, "Python id": 12910, "finished": true, "Ev Idx": 15810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method zero_ of Tensor object at 0x7f6dc8017bb0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917367.672, "dur": 8.647,
    "args": {
      "Python parent id": 12, "Python id": 12911, "finished": true, "Ev Idx": 15811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7f6dadbb4180>", "pid": 49625, "tid": 49625,
    "ts": 7253740917376.813, "dur": 0.099,
    "args": {
      "Python parent id": 12, "Python id": 12912, "finished": true, "Ev Idx": 15812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917377.637, "dur": 0.218,
    "args": {
      "Python parent id": 12, "Python id": 12913, "finished": true, "Ev Idx": 15813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917378.360, "dur": 0.085,
    "args": {
      "Python parent id": 12, "Python id": 12914, "finished": true, "Ev Idx": 15814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7f78ea90c730>", "pid": 49625, "tid": 49625,
    "ts": 7253740917378.858, "dur": 10.615,
    "args": {
      "Python parent id": 12, "Python id": 12915, "finished": true, "Ev Idx": 15815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917390.856, "dur": 15.764,
    "args": {
      "Python parent id": 12, "Python id": 12916, "finished": true, "Ev Idx": 15816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917391.231, "dur": 11.908,
    "args": {
      "Python parent id": 12916, "Python id": 12917, "finished": true, "Ev Idx": 15817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917391.383, "dur": 0.532,
    "args": {
      "Python parent id": 12917, "Python id": 12918, "finished": true, "Ev Idx": 15818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917391.550, "dur": 0.275,
    "args": {
      "Python parent id": 12918, "Python id": 12919, "finished": true, "Ev Idx": 15819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917391.659, "dur": 0.109,
    "args": {
      "Python parent id": 12919, "Python id": 12920, "finished": true, "Ev Idx": 15820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917392.181, "dur": 8.941,
    "args": {
      "Python parent id": 12917, "Python id": 12921, "finished": true, "Ev Idx": 15821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917392.268, "dur": 0.071,
    "args": {
      "Python parent id": 12921, "Python id": 12922, "finished": true, "Ev Idx": 15822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917392.481, "dur": 0.063,
    "args": {
      "Python parent id": 12921, "Python id": 12923, "finished": true, "Ev Idx": 15823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917392.713, "dur": 0.052,
    "args": {
      "Python parent id": 12921, "Python id": 12924, "finished": true, "Ev Idx": 15824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917393.236, "dur": 0.044,
    "args": {
      "Python parent id": 12921, "Python id": 12925, "finished": true, "Ev Idx": 15825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917393.536, "dur": 0.079,
    "args": {
      "Python parent id": 12921, "Python id": 12926, "finished": true, "Ev Idx": 15826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917393.937, "dur": 7.102,
    "args": {
      "Python parent id": 12921, "Python id": 12927, "finished": true, "Ev Idx": 15827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917394.034, "dur": 0.059,
    "args": {
      "Python parent id": 12927, "Python id": 12928, "finished": true, "Ev Idx": 15828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917394.259, "dur": 0.061,
    "args": {
      "Python parent id": 12927, "Python id": 12929, "finished": true, "Ev Idx": 15829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917394.477, "dur": 0.069,
    "args": {
      "Python parent id": 12927, "Python id": 12930, "finished": true, "Ev Idx": 15830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917394.925, "dur": 0.076,
    "args": {
      "Python parent id": 12927, "Python id": 12931, "finished": true, "Ev Idx": 15831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917395.324, "dur": 5.593,
    "args": {
      "Python parent id": 12927, "Python id": 12932, "finished": true, "Ev Idx": 15832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917395.557, "dur": 5.238,
    "args": {
      "Python parent id": 12932, "Python id": 12933, "finished": true, "Ev Idx": 15833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917395.690, "dur": 3.353,
    "args": {
      "Python parent id": 12933, "Python id": 12934, "finished": true, "Ev Idx": 15834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917395.894, "dur": 3.081,
    "args": {
      "Python parent id": 12934, "Python id": 12935, "finished": true, "Ev Idx": 15835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917396.068, "dur": 0.292,
    "args": {
      "Python parent id": 12935, "Python id": 12936, "finished": true, "Ev Idx": 15836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917396.201, "dur": 0.085,
    "args": {
      "Python parent id": 12936, "Python id": 12937, "finished": true, "Ev Idx": 15837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917396.522, "dur": 2.105,
    "args": {
      "Python parent id": 12935, "Python id": 12938, "finished": true, "Ev Idx": 15838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917396.730, "dur": 1.814,
    "args": {
      "Python parent id": 12938, "Python id": 12939, "finished": true, "Ev Idx": 15839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917396.920, "dur": 1.561,
    "args": {
      "Python parent id": 12939, "Python id": 12940, "finished": true, "Ev Idx": 15840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917397.068, "dur": 1.199,
    "args": {
      "Python parent id": 12940, "Python id": 12941, "finished": true, "Ev Idx": 15841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917397.277, "dur": 0.556,
    "args": {
      "Python parent id": 12941, "Python id": 12942, "finished": true, "Ev Idx": 15842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917397.381, "dur": 0.072,
    "args": {
      "Python parent id": 12942, "Python id": 12943, "finished": true, "Ev Idx": 15843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917397.615, "dur": 0.137,
    "args": {
      "Python parent id": 12942, "Python id": 12944, "finished": true, "Ev Idx": 15844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917398.822, "dur": 0.105,
    "args": {
      "Python parent id": 12935, "Python id": 12945, "finished": true, "Ev Idx": 15845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917399.226, "dur": 0.083,
    "args": {
      "Python parent id": 12933, "Python id": 12946, "finished": true, "Ev Idx": 15846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917399.593, "dur": 1.137,
    "args": {
      "Python parent id": 12933, "Python id": 12947, "finished": true, "Ev Idx": 15847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917399.767, "dur": 0.922,
    "args": {
      "Python parent id": 12947, "Python id": 12948, "finished": true, "Ev Idx": 15848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917399.905, "dur": 0.375,
    "args": {
      "Python parent id": 12948, "Python id": 12949, "finished": true, "Ev Idx": 15849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917400.043, "dur": 0.178,
    "args": {
      "Python parent id": 12949, "Python id": 12950, "finished": true, "Ev Idx": 15850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917400.136, "dur": 0.055,
    "args": {
      "Python parent id": 12950, "Python id": 12951, "finished": true, "Ev Idx": 15851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917400.417, "dur": 0.234,
    "args": {
      "Python parent id": 12948, "Python id": 12952, "finished": true, "Ev Idx": 15852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917401.243, "dur": 0.185,
    "args": {
      "Python parent id": 12917, "Python id": 12953, "finished": true, "Ev Idx": 15853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917402.007, "dur": 1.006,
    "args": {
      "Python parent id": 12917, "Python id": 12954, "finished": true, "Ev Idx": 15854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917402.439, "dur": 0.520,
    "args": {
      "Python parent id": 12954, "Python id": 12955, "finished": true, "Ev Idx": 15855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917403.409, "dur": 3.159,
    "args": {
      "Python parent id": 12916, "Python id": 12956, "finished": true, "Ev Idx": 15856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "bench_qdq_all.py(10): run_scaled_fake_quantize_eager", "pid": 49625, "tid": 49625,
    "ts": 7253740917407.597, "dur": 93.439,
    "args": {
      "Python parent id": 12, "Python id": 12957, "finished": true, "Ev Idx": 15857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917407.989, "dur": 0.474,
    "args": {
      "Python parent id": 12957, "Python id": 12958, "finished": true, "Ev Idx": 15858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917408.308, "dur": 0.081,
    "args": {
      "Python parent id": 12958, "Python id": 12959, "finished": true, "Ev Idx": 15859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917408.743, "dur": 0.399,
    "args": {
      "Python parent id": 12957, "Python id": 12960, "finished": true, "Ev Idx": 15860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 49625, "tid": 49625,
    "ts": 7253740917409.354, "dur": 0.270,
    "args": {
      "Python parent id": 12957, "Python id": 12961, "finished": true, "Ev Idx": 15861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7f6dada20f50>", "pid": 49625, "tid": 49625,
    "ts": 7253740917409.847, "dur": 1.693,
    "args": {
      "Python parent id": 12957, "Python id": 12962, "finished": true, "Ev Idx": 15862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/__init__.py(118): _scaled_fake_quantize_forward", "pid": 49625, "tid": 49625,
    "ts": 7253740917412.118, "dur": 88.819,
    "args": {
      "Python parent id": 12957, "Python id": 12963, "finished": true, "Ev Idx": 15863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_ops.py(1147): __call__", "pid": 49625, "tid": 49625,
    "ts": 7253740917413.038, "dur": 87.694,
    "args": {
      "Python parent id": 12963, "Python id": 12964, "finished": true, "Ev Idx": 15864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method scaled_fake_quantize of PyCapsule object at 0x7f76a2cf13b0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917413.374, "dur": 87.284,
    "args": {
      "Python parent id": 12964, "Python id": 12965, "finished": true, "Ev Idx": 15865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(175): scaled_fake_quantize", "pid": 49625, "tid": 49625,
    "ts": 7253740917418.769, "dur": 80.470,
    "args": {
      "Python parent id": 12965, "Python id": 12966, "finished": true, "Ev Idx": 15866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/hw_emulation_interface.py(183): scaled_fake_quantize_impl", "pid": 49625, "tid": 49625,
    "ts": 7253740917419.096, "dur": 80.037,
    "args": {
      "Python parent id": 12966, "Python id": 12967, "finished": true, "Ev Idx": 15867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(36): fake_quantize_int", "pid": 49625, "tid": 49625,
    "ts": 7253740917419.745, "dur": 79.205,
    "args": {
      "Python parent id": 12967, "Python id": 12968, "finished": true, "Ev Idx": 15868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917420.261, "dur": 0.441,
    "args": {
      "Python parent id": 12968, "Python id": 12969, "finished": true, "Ev Idx": 15869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917420.561, "dur": 0.060,
    "args": {
      "Python parent id": 12969, "Python id": 12970, "finished": true, "Ev Idx": 15870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(202): __get__", "pid": 49625, "tid": 49625,
    "ts": 7253740917421.038, "dur": 0.358,
    "args": {
      "Python parent id": 12968, "Python id": 12971, "finished": true, "Ev Idx": 15871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1286): value", "pid": 49625, "tid": 49625,
    "ts": 7253740917421.289, "dur": 0.052,
    "args": {
      "Python parent id": 12971, "Python id": 12972, "finished": true, "Ev Idx": 15872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(394): fake_quantize_int_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917421.749, "dur": 77.092,
    "args": {
      "Python parent id": 12968, "Python id": 12973, "finished": true, "Ev Idx": 15873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "quark/torch/kernel/hw_emulation/fake_quantize.py(373): fake_quantize_per_channel_affine", "pid": 49625, "tid": 49625,
    "ts": 7253740917422.493, "dur": 75.042,
    "args": {
      "Python parent id": 12973, "Python id": 12974, "finished": true, "Ev Idx": 15874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(33): wrapped", "pid": 49625, "tid": 49625,
    "ts": 7253740917429.839, "dur": 19.685,
    "args": {
      "Python parent id": 12974, "Python id": 12975, "finished": true, "Ev Idx": 15875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function>", "pid": 49625, "tid": 49625,
    "ts": 7253740917429.986, "dur": 0.165,
    "args": {
      "Python parent id": 12975, "Python id": 12976, "finished": true, "Ev Idx": 15876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_tensor.py(1100): __rdiv__", "pid": 49625, "tid": 49625,
    "ts": 7253740917430.454, "dur": 18.920,
    "args": {
      "Python parent id": 12975, "Python id": 12977, "finished": true, "Ev Idx": 15877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reciprocal of Tensor object at 0x7f6dada21bd0>", "pid": 49625, "tid": 49625,
    "ts": 7253740917430.601, "dur": 9.133,
    "args": {
      "Python parent id": 12977, "Python id": 12978, "finished": true, "Ev Idx": 15878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method round of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917457.436, "dur": 7.408,
    "args": {
      "Python parent id": 12974, "Python id": 12979, "finished": true, "Ev Idx": 15879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clamp of type object at 0x7f78e9d5bf40>", "pid": 49625, "tid": 49625,
    "ts": 7253740917473.455, "dur": 9.021,
    "args": {
      "Python parent id": 12974, "Python id": 12980, "finished": true, "Ev Idx": 15880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(179): record", "pid": 49625, "tid": 49625,
    "ts": 7253740917502.454, "dur": 17.219,
    "args": {
      "Python parent id": 12, "Python id": 12981, "finished": true, "Ev Idx": 15881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1056): current_stream", "pid": 49625, "tid": 49625,
    "ts": 7253740917502.927, "dur": 13.047,
    "args": {
      "Python parent id": 12981, "Python id": 12982, "finished": true, "Ev Idx": 15882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917503.139, "dur": 0.446,
    "args": {
      "Python parent id": 12982, "Python id": 12983, "finished": true, "Ev Idx": 15883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917503.291, "dur": 0.229,
    "args": {
      "Python parent id": 12983, "Python id": 12984, "finished": true, "Ev Idx": 15884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917503.394, "dur": 0.088,
    "args": {
      "Python parent id": 12984, "Python id": 12985, "finished": true, "Ev Idx": 15885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917503.923, "dur": 9.596,
    "args": {
      "Python parent id": 12982, "Python id": 12986, "finished": true, "Ev Idx": 15886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917504.043, "dur": 0.088,
    "args": {
      "Python parent id": 12986, "Python id": 12987, "finished": true, "Ev Idx": 15887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917504.245, "dur": 0.068,
    "args": {
      "Python parent id": 12986, "Python id": 12988, "finished": true, "Ev Idx": 15888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917504.463, "dur": 0.062,
    "args": {
      "Python parent id": 12986, "Python id": 12989, "finished": true, "Ev Idx": 15889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917505.032, "dur": 0.030,
    "args": {
      "Python parent id": 12986, "Python id": 12990, "finished": true, "Ev Idx": 15890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917505.400, "dur": 0.076,
    "args": {
      "Python parent id": 12986, "Python id": 12991, "finished": true, "Ev Idx": 15891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917505.812, "dur": 7.633,
    "args": {
      "Python parent id": 12986, "Python id": 12992, "finished": true, "Ev Idx": 15892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917505.910, "dur": 0.069,
    "args": {
      "Python parent id": 12992, "Python id": 12993, "finished": true, "Ev Idx": 15893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917506.155, "dur": 0.062,
    "args": {
      "Python parent id": 12992, "Python id": 12994, "finished": true, "Ev Idx": 15894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917506.405, "dur": 0.070,
    "args": {
      "Python parent id": 12992, "Python id": 12995, "finished": true, "Ev Idx": 15895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917506.806, "dur": 0.041,
    "args": {
      "Python parent id": 12992, "Python id": 12996, "finished": true, "Ev Idx": 15896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917507.156, "dur": 6.160,
    "args": {
      "Python parent id": 12992, "Python id": 12997, "finished": true, "Ev Idx": 15897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917507.443, "dur": 5.777,
    "args": {
      "Python parent id": 12997, "Python id": 12998, "finished": true, "Ev Idx": 15898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917507.568, "dur": 3.708,
    "args": {
      "Python parent id": 12998, "Python id": 12999, "finished": true, "Ev Idx": 15899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917507.742, "dur": 3.429,
    "args": {
      "Python parent id": 12999, "Python id": 13000, "finished": true, "Ev Idx": 15900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917507.888, "dur": 0.313,
    "args": {
      "Python parent id": 13000, "Python id": 13001, "finished": true, "Ev Idx": 15901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917508.025, "dur": 0.101,
    "args": {
      "Python parent id": 13001, "Python id": 13002, "finished": true, "Ev Idx": 15902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917508.399, "dur": 2.347,
    "args": {
      "Python parent id": 13000, "Python id": 13003, "finished": true, "Ev Idx": 15903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917508.620, "dur": 2.012,
    "args": {
      "Python parent id": 13003, "Python id": 13004, "finished": true, "Ev Idx": 15904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917508.833, "dur": 1.728,
    "args": {
      "Python parent id": 13004, "Python id": 13005, "finished": true, "Ev Idx": 15905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917509.036, "dur": 1.296,
    "args": {
      "Python parent id": 13005, "Python id": 13006, "finished": true, "Ev Idx": 15906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917509.227, "dur": 0.554,
    "args": {
      "Python parent id": 13006, "Python id": 13007, "finished": true, "Ev Idx": 15907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917509.314, "dur": 0.067,
    "args": {
      "Python parent id": 13007, "Python id": 13008, "finished": true, "Ev Idx": 15908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917509.547, "dur": 0.157,
    "args": {
      "Python parent id": 13007, "Python id": 13009, "finished": true, "Ev Idx": 15909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917510.936, "dur": 0.177,
    "args": {
      "Python parent id": 13000, "Python id": 13010, "finished": true, "Ev Idx": 15910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917511.463, "dur": 0.103,
    "args": {
      "Python parent id": 12998, "Python id": 13011, "finished": true, "Ev Idx": 15911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917511.933, "dur": 1.224,
    "args": {
      "Python parent id": 12998, "Python id": 13012, "finished": true, "Ev Idx": 15912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917512.082, "dur": 1.036,
    "args": {
      "Python parent id": 13012, "Python id": 13013, "finished": true, "Ev Idx": 15913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917512.230, "dur": 0.383,
    "args": {
      "Python parent id": 13013, "Python id": 13014, "finished": true, "Ev Idx": 15914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917512.368, "dur": 0.183,
    "args": {
      "Python parent id": 13014, "Python id": 13015, "finished": true, "Ev Idx": 15915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917512.461, "dur": 0.057,
    "args": {
      "Python parent id": 13015, "Python id": 13016, "finished": true, "Ev Idx": 15916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917512.741, "dur": 0.312,
    "args": {
      "Python parent id": 13013, "Python id": 13017, "finished": true, "Ev Idx": 15917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getCurrentStream>", "pid": 49625, "tid": 49625,
    "ts": 7253740917513.659, "dur": 0.349,
    "args": {
      "Python parent id": 12982, "Python id": 13018, "finished": true, "Ev Idx": 15918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(34): __new__", "pid": 49625, "tid": 49625,
    "ts": 7253740917514.688, "dur": 1.114,
    "args": {
      "Python parent id": 12982, "Python id": 13019, "finished": true, "Ev Idx": 15919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x7f78e9d5f200>", "pid": 49625, "tid": 49625,
    "ts": 7253740917515.221, "dur": 0.529,
    "args": {
      "Python parent id": 13019, "Python id": 13020, "finished": true, "Ev Idx": 15920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method record of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740917516.304, "dur": 3.280,
    "args": {
      "Python parent id": 12981, "Python id": 13021, "finished": true, "Ev Idx": 15921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1030): synchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740917521.927, "dur": 43298.071,
    "args": {
      "Python parent id": 12, "Python id": 13022, "finished": true, "Ev Idx": 15922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917522.130, "dur": 0.466,
    "args": {
      "Python parent id": 13022, "Python id": 13023, "finished": true, "Ev Idx": 15923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917522.311, "dur": 0.198,
    "args": {
      "Python parent id": 13023, "Python id": 13024, "finished": true, "Ev Idx": 15924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917522.409, "dur": 0.064,
    "args": {
      "Python parent id": 13024, "Python id": 13025, "finished": true, "Ev Idx": 15925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(490): __init__", "pid": 49625, "tid": 49625,
    "ts": 7253740917524.118, "dur": 9.353,
    "args": {
      "Python parent id": 13022, "Python id": 13026, "finished": true, "Ev Idx": 15926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917524.493, "dur": 8.775,
    "args": {
      "Python parent id": 13026, "Python id": 13027, "finished": true, "Ev Idx": 15927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917524.598, "dur": 0.074,
    "args": {
      "Python parent id": 13027, "Python id": 13028, "finished": true, "Ev Idx": 15928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917524.804, "dur": 0.062,
    "args": {
      "Python parent id": 13027, "Python id": 13029, "finished": true, "Ev Idx": 15929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917525.029, "dur": 0.062,
    "args": {
      "Python parent id": 13027, "Python id": 13030, "finished": true, "Ev Idx": 15930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917525.438, "dur": 0.040,
    "args": {
      "Python parent id": 13027, "Python id": 13031, "finished": true, "Ev Idx": 15931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917525.765, "dur": 0.081,
    "args": {
      "Python parent id": 13027, "Python id": 13032, "finished": true, "Ev Idx": 15932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917526.246, "dur": 6.872,
    "args": {
      "Python parent id": 13027, "Python id": 13033, "finished": true, "Ev Idx": 15933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917526.325, "dur": 0.061,
    "args": {
      "Python parent id": 13033, "Python id": 13034, "finished": true, "Ev Idx": 15934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917526.537, "dur": 0.064,
    "args": {
      "Python parent id": 13033, "Python id": 13035, "finished": true, "Ev Idx": 15935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917526.778, "dur": 0.069,
    "args": {
      "Python parent id": 13033, "Python id": 13036, "finished": true, "Ev Idx": 15936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740917527.193, "dur": 0.045,
    "args": {
      "Python parent id": 13033, "Python id": 13037, "finished": true, "Ev Idx": 15937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740917527.578, "dur": 5.480,
    "args": {
      "Python parent id": 13033, "Python id": 13038, "finished": true, "Ev Idx": 15938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740917527.769, "dur": 5.225,
    "args": {
      "Python parent id": 13038, "Python id": 13039, "finished": true, "Ev Idx": 15939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740917527.902, "dur": 3.324,
    "args": {
      "Python parent id": 13039, "Python id": 13040, "finished": true, "Ev Idx": 15940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740917528.111, "dur": 3.047,
    "args": {
      "Python parent id": 13040, "Python id": 13041, "finished": true, "Ev Idx": 15941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740917528.244, "dur": 0.254,
    "args": {
      "Python parent id": 13041, "Python id": 13042, "finished": true, "Ev Idx": 15942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740917528.367, "dur": 0.072,
    "args": {
      "Python parent id": 13042, "Python id": 13043, "finished": true, "Ev Idx": 15943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740917528.690, "dur": 2.143,
    "args": {
      "Python parent id": 13041, "Python id": 13044, "finished": true, "Ev Idx": 15944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740917528.859, "dur": 1.916,
    "args": {
      "Python parent id": 13044, "Python id": 13045, "finished": true, "Ev Idx": 15945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740917529.062, "dur": 1.663,
    "args": {
      "Python parent id": 13045, "Python id": 13046, "finished": true, "Ev Idx": 15946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740917529.228, "dur": 1.222,
    "args": {
      "Python parent id": 13046, "Python id": 13047, "finished": true, "Ev Idx": 15947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740917529.444, "dur": 0.530,
    "args": {
      "Python parent id": 13047, "Python id": 13048, "finished": true, "Ev Idx": 15948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740917529.530, "dur": 0.069,
    "args": {
      "Python parent id": 13048, "Python id": 13049, "finished": true, "Ev Idx": 15949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740917529.797, "dur": 0.137,
    "args": {
      "Python parent id": 13048, "Python id": 13050, "finished": true, "Ev Idx": 15950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740917531.027, "dur": 0.099,
    "args": {
      "Python parent id": 13041, "Python id": 13051, "finished": true, "Ev Idx": 15951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740917531.393, "dur": 0.113,
    "args": {
      "Python parent id": 13039, "Python id": 13052, "finished": true, "Ev Idx": 15952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740917531.781, "dur": 1.146,
    "args": {
      "Python parent id": 13039, "Python id": 13053, "finished": true, "Ev Idx": 15953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740917531.973, "dur": 0.909,
    "args": {
      "Python parent id": 13053, "Python id": 13054, "finished": true, "Ev Idx": 15954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740917532.146, "dur": 0.349,
    "args": {
      "Python parent id": 13054, "Python id": 13055, "finished": true, "Ev Idx": 15955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740917532.270, "dur": 0.168,
    "args": {
      "Python parent id": 13055, "Python id": 13056, "finished": true, "Ev Idx": 15956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740917532.357, "dur": 0.049,
    "args": {
      "Python parent id": 13056, "Python id": 13057, "finished": true, "Ev Idx": 15957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917532.637, "dur": 0.216,
    "args": {
      "Python parent id": 13054, "Python id": 13058, "finished": true, "Ev Idx": 15958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(494): __enter__", "pid": 49625, "tid": 49625,
    "ts": 7253740917534.482, "dur": 1.218,
    "args": {
      "Python parent id": 13022, "Python id": 13059, "finished": true, "Ev Idx": 15959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_exchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740917534.994, "dur": 0.558,
    "args": {
      "Python parent id": 13059, "Python id": 13060, "finished": true, "Ev Idx": 15960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_synchronize>", "pid": 49625, "tid": 49625,
    "ts": 7253740917536.361, "dur": 43280.766,
    "args": {
      "Python parent id": 13022, "Python id": 13061, "finished": true, "Ev Idx": 15961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(497): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740960818.156, "dur": 1.314,
    "args": {
      "Python parent id": 13022, "Python id": 13062, "finished": true, "Ev Idx": 15962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_maybeExchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740960818.606, "dur": 0.760,
    "args": {
      "Python parent id": 13062, "Python id": 13063, "finished": true, "Ev Idx": 15963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960825.088, "dur": 40.610,
    "args": {
      "Python parent id": 12, "Python id": 13064, "finished": true, "Ev Idx": 15964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960861.069, "dur": 4.488,
    "args": {
      "Python parent id": 13064, "Python id": 13065, "finished": true, "Ev Idx": 15965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960867.472, "dur": 2.241,
    "args": {
      "Python parent id": 12, "Python id": 13066, "finished": true, "Ev Idx": 15966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960868.210, "dur": 1.453,
    "args": {
      "Python parent id": 13066, "Python id": 13067, "finished": true, "Ev Idx": 15967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960871.472, "dur": 2.543,
    "args": {
      "Python parent id": 12, "Python id": 13068, "finished": true, "Ev Idx": 15968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960871.995, "dur": 1.858,
    "args": {
      "Python parent id": 13068, "Python id": 13069, "finished": true, "Ev Idx": 15969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960875.920, "dur": 1.621,
    "args": {
      "Python parent id": 12, "Python id": 13070, "finished": true, "Ev Idx": 15970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960876.197, "dur": 1.023,
    "args": {
      "Python parent id": 13070, "Python id": 13071, "finished": true, "Ev Idx": 15971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960878.390, "dur": 1.432,
    "args": {
      "Python parent id": 12, "Python id": 13072, "finished": true, "Ev Idx": 15972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960878.591, "dur": 0.952,
    "args": {
      "Python parent id": 13072, "Python id": 13073, "finished": true, "Ev Idx": 15973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960880.665, "dur": 1.406,
    "args": {
      "Python parent id": 12, "Python id": 13074, "finished": true, "Ev Idx": 15974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960881.091, "dur": 0.910,
    "args": {
      "Python parent id": 13074, "Python id": 13075, "finished": true, "Ev Idx": 15975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960882.702, "dur": 1.510,
    "args": {
      "Python parent id": 12, "Python id": 13076, "finished": true, "Ev Idx": 15976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960882.964, "dur": 1.081,
    "args": {
      "Python parent id": 13076, "Python id": 13077, "finished": true, "Ev Idx": 15977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960884.816, "dur": 1.117,
    "args": {
      "Python parent id": 12, "Python id": 13078, "finished": true, "Ev Idx": 15978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960885.016, "dur": 0.872,
    "args": {
      "Python parent id": 13078, "Python id": 13079, "finished": true, "Ev Idx": 15979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960886.579, "dur": 1.230,
    "args": {
      "Python parent id": 12, "Python id": 13080, "finished": true, "Ev Idx": 15980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960886.809, "dur": 0.951,
    "args": {
      "Python parent id": 13080, "Python id": 13081, "finished": true, "Ev Idx": 15981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960888.608, "dur": 1.895,
    "args": {
      "Python parent id": 12, "Python id": 13082, "finished": true, "Ev Idx": 15982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960888.872, "dur": 1.316,
    "args": {
      "Python parent id": 13082, "Python id": 13083, "finished": true, "Ev Idx": 15983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960891.149, "dur": 1.467,
    "args": {
      "Python parent id": 12, "Python id": 13084, "finished": true, "Ev Idx": 15984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960891.366, "dur": 0.956,
    "args": {
      "Python parent id": 13084, "Python id": 13085, "finished": true, "Ev Idx": 15985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960893.221, "dur": 1.313,
    "args": {
      "Python parent id": 12, "Python id": 13086, "finished": true, "Ev Idx": 15986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960893.440, "dur": 1.018,
    "args": {
      "Python parent id": 13086, "Python id": 13087, "finished": true, "Ev Idx": 15987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960895.108, "dur": 1.220,
    "args": {
      "Python parent id": 12, "Python id": 13088, "finished": true, "Ev Idx": 15988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960895.304, "dur": 0.945,
    "args": {
      "Python parent id": 13088, "Python id": 13089, "finished": true, "Ev Idx": 15989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960896.911, "dur": 1.188,
    "args": {
      "Python parent id": 12, "Python id": 13090, "finished": true, "Ev Idx": 15990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960897.143, "dur": 0.912,
    "args": {
      "Python parent id": 13090, "Python id": 13091, "finished": true, "Ev Idx": 15991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960898.801, "dur": 1.322,
    "args": {
      "Python parent id": 12, "Python id": 13092, "finished": true, "Ev Idx": 15992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960898.995, "dur": 1.077,
    "args": {
      "Python parent id": 13092, "Python id": 13093, "finished": true, "Ev Idx": 15993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960900.820, "dur": 1.316,
    "args": {
      "Python parent id": 12, "Python id": 13094, "finished": true, "Ev Idx": 15994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960901.049, "dur": 0.955,
    "args": {
      "Python parent id": 13094, "Python id": 13095, "finished": true, "Ev Idx": 15995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960902.753, "dur": 1.290,
    "args": {
      "Python parent id": 12, "Python id": 13096, "finished": true, "Ev Idx": 15996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960902.948, "dur": 0.938,
    "args": {
      "Python parent id": 13096, "Python id": 13097, "finished": true, "Ev Idx": 15997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960904.820, "dur": 1.126,
    "args": {
      "Python parent id": 12, "Python id": 13098, "finished": true, "Ev Idx": 15998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960905.017, "dur": 0.879,
    "args": {
      "Python parent id": 13098, "Python id": 13099, "finished": true, "Ev Idx": 15999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960906.598, "dur": 1.462,
    "args": {
      "Python parent id": 12, "Python id": 13100, "finished": true, "Ev Idx": 16000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960906.829, "dur": 1.101,
    "args": {
      "Python parent id": 13100, "Python id": 13101, "finished": true, "Ev Idx": 16001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960908.703, "dur": 1.358,
    "args": {
      "Python parent id": 12, "Python id": 13102, "finished": true, "Ev Idx": 16002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960908.900, "dur": 1.109,
    "args": {
      "Python parent id": 13102, "Python id": 13103, "finished": true, "Ev Idx": 16003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960910.766, "dur": 1.325,
    "args": {
      "Python parent id": 12, "Python id": 13104, "finished": true, "Ev Idx": 16004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960910.960, "dur": 0.989,
    "args": {
      "Python parent id": 13104, "Python id": 13105, "finished": true, "Ev Idx": 16005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960912.690, "dur": 1.139,
    "args": {
      "Python parent id": 12, "Python id": 13106, "finished": true, "Ev Idx": 16006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960912.888, "dur": 0.888,
    "args": {
      "Python parent id": 13106, "Python id": 13107, "finished": true, "Ev Idx": 16007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960914.408, "dur": 1.162,
    "args": {
      "Python parent id": 12, "Python id": 13108, "finished": true, "Ev Idx": 16008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960914.606, "dur": 0.788,
    "args": {
      "Python parent id": 13108, "Python id": 13109, "finished": true, "Ev Idx": 16009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960916.197, "dur": 1.302,
    "args": {
      "Python parent id": 12, "Python id": 13110, "finished": true, "Ev Idx": 16010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960916.393, "dur": 1.058,
    "args": {
      "Python parent id": 13110, "Python id": 13111, "finished": true, "Ev Idx": 16011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960918.202, "dur": 1.261,
    "args": {
      "Python parent id": 12, "Python id": 13112, "finished": true, "Ev Idx": 16012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960918.436, "dur": 0.861,
    "args": {
      "Python parent id": 13112, "Python id": 13113, "finished": true, "Ev Idx": 16013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960920.119, "dur": 1.282,
    "args": {
      "Python parent id": 12, "Python id": 13114, "finished": true, "Ev Idx": 16014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960920.313, "dur": 0.937,
    "args": {
      "Python parent id": 13114, "Python id": 13115, "finished": true, "Ev Idx": 16015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960922.029, "dur": 1.381,
    "args": {
      "Python parent id": 12, "Python id": 13116, "finished": true, "Ev Idx": 16016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960922.223, "dur": 1.136,
    "args": {
      "Python parent id": 13116, "Python id": 13117, "finished": true, "Ev Idx": 16017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960923.993, "dur": 1.157,
    "args": {
      "Python parent id": 12, "Python id": 13118, "finished": true, "Ev Idx": 16018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960924.190, "dur": 0.909,
    "args": {
      "Python parent id": 13118, "Python id": 13119, "finished": true, "Ev Idx": 16019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960925.839, "dur": 1.381,
    "args": {
      "Python parent id": 12, "Python id": 13120, "finished": true, "Ev Idx": 16020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960926.034, "dur": 0.941,
    "args": {
      "Python parent id": 13120, "Python id": 13121, "finished": true, "Ev Idx": 16021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960927.799, "dur": 1.608,
    "args": {
      "Python parent id": 12, "Python id": 13122, "finished": true, "Ev Idx": 16022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960927.996, "dur": 1.091,
    "args": {
      "Python parent id": 13122, "Python id": 13123, "finished": true, "Ev Idx": 16023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960930.037, "dur": 1.162,
    "args": {
      "Python parent id": 12, "Python id": 13124, "finished": true, "Ev Idx": 16024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960930.233, "dur": 0.910,
    "args": {
      "Python parent id": 13124, "Python id": 13125, "finished": true, "Ev Idx": 16025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960931.809, "dur": 1.669,
    "args": {
      "Python parent id": 12, "Python id": 13126, "finished": true, "Ev Idx": 16026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960932.040, "dur": 1.389,
    "args": {
      "Python parent id": 13126, "Python id": 13127, "finished": true, "Ev Idx": 16027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960934.074, "dur": 1.326,
    "args": {
      "Python parent id": 12, "Python id": 13128, "finished": true, "Ev Idx": 16028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960934.268, "dur": 1.088,
    "args": {
      "Python parent id": 13128, "Python id": 13129, "finished": true, "Ev Idx": 16029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960936.545, "dur": 1.236,
    "args": {
      "Python parent id": 12, "Python id": 13130, "finished": true, "Ev Idx": 16030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960936.743, "dur": 0.988,
    "args": {
      "Python parent id": 13130, "Python id": 13131, "finished": true, "Ev Idx": 16031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960938.629, "dur": 1.254,
    "args": {
      "Python parent id": 12, "Python id": 13132, "finished": true, "Ev Idx": 16032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960938.825, "dur": 1.004,
    "args": {
      "Python parent id": 13132, "Python id": 13133, "finished": true, "Ev Idx": 16033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960940.500, "dur": 1.131,
    "args": {
      "Python parent id": 12, "Python id": 13134, "finished": true, "Ev Idx": 16034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960940.698, "dur": 0.881,
    "args": {
      "Python parent id": 13134, "Python id": 13135, "finished": true, "Ev Idx": 16035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960942.318, "dur": 1.288,
    "args": {
      "Python parent id": 12, "Python id": 13136, "finished": true, "Ev Idx": 16036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960942.520, "dur": 1.037,
    "args": {
      "Python parent id": 13136, "Python id": 13137, "finished": true, "Ev Idx": 16037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960944.188, "dur": 1.639,
    "args": {
      "Python parent id": 12, "Python id": 13138, "finished": true, "Ev Idx": 16038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960944.392, "dur": 1.385,
    "args": {
      "Python parent id": 13138, "Python id": 13139, "finished": true, "Ev Idx": 16039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960946.530, "dur": 1.256,
    "args": {
      "Python parent id": 12, "Python id": 13140, "finished": true, "Ev Idx": 16040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960946.723, "dur": 1.011,
    "args": {
      "Python parent id": 13140, "Python id": 13141, "finished": true, "Ev Idx": 16041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960948.381, "dur": 1.183,
    "args": {
      "Python parent id": 12, "Python id": 13142, "finished": true, "Ev Idx": 16042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960948.627, "dur": 0.886,
    "args": {
      "Python parent id": 13142, "Python id": 13143, "finished": true, "Ev Idx": 16043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960950.245, "dur": 1.495,
    "args": {
      "Python parent id": 12, "Python id": 13144, "finished": true, "Ev Idx": 16044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960950.438, "dur": 1.255,
    "args": {
      "Python parent id": 13144, "Python id": 13145, "finished": true, "Ev Idx": 16045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960952.631, "dur": 1.418,
    "args": {
      "Python parent id": 12, "Python id": 13146, "finished": true, "Ev Idx": 16046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960952.827, "dur": 1.169,
    "args": {
      "Python parent id": 13146, "Python id": 13147, "finished": true, "Ev Idx": 16047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960954.756, "dur": 1.163,
    "args": {
      "Python parent id": 12, "Python id": 13148, "finished": true, "Ev Idx": 16048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960954.950, "dur": 0.918,
    "args": {
      "Python parent id": 13148, "Python id": 13149, "finished": true, "Ev Idx": 16049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960956.657, "dur": 1.308,
    "args": {
      "Python parent id": 12, "Python id": 13150, "finished": true, "Ev Idx": 16050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960956.858, "dur": 1.055,
    "args": {
      "Python parent id": 13150, "Python id": 13151, "finished": true, "Ev Idx": 16051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960958.766, "dur": 1.131,
    "args": {
      "Python parent id": 12, "Python id": 13152, "finished": true, "Ev Idx": 16052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960958.959, "dur": 0.890,
    "args": {
      "Python parent id": 13152, "Python id": 13153, "finished": true, "Ev Idx": 16053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960960.724, "dur": 1.435,
    "args": {
      "Python parent id": 12, "Python id": 13154, "finished": true, "Ev Idx": 16054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960960.923, "dur": 1.185,
    "args": {
      "Python parent id": 13154, "Python id": 13155, "finished": true, "Ev Idx": 16055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960962.897, "dur": 1.668,
    "args": {
      "Python parent id": 12, "Python id": 13156, "finished": true, "Ev Idx": 16056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960963.093, "dur": 1.422,
    "args": {
      "Python parent id": 13156, "Python id": 13157, "finished": true, "Ev Idx": 16057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960965.141, "dur": 1.435,
    "args": {
      "Python parent id": 12, "Python id": 13158, "finished": true, "Ev Idx": 16058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960965.339, "dur": 1.187,
    "args": {
      "Python parent id": 13158, "Python id": 13159, "finished": true, "Ev Idx": 16059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960967.288, "dur": 1.211,
    "args": {
      "Python parent id": 12, "Python id": 13160, "finished": true, "Ev Idx": 16060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960967.513, "dur": 0.933,
    "args": {
      "Python parent id": 13160, "Python id": 13161, "finished": true, "Ev Idx": 16061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960969.103, "dur": 1.075,
    "args": {
      "Python parent id": 12, "Python id": 13162, "finished": true, "Ev Idx": 16062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960969.300, "dur": 0.831,
    "args": {
      "Python parent id": 13162, "Python id": 13163, "finished": true, "Ev Idx": 16063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960970.883, "dur": 1.107,
    "args": {
      "Python parent id": 12, "Python id": 13164, "finished": true, "Ev Idx": 16064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960971.076, "dur": 0.869,
    "args": {
      "Python parent id": 13164, "Python id": 13165, "finished": true, "Ev Idx": 16065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960972.704, "dur": 1.125,
    "args": {
      "Python parent id": 12, "Python id": 13166, "finished": true, "Ev Idx": 16066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960972.907, "dur": 0.876,
    "args": {
      "Python parent id": 13166, "Python id": 13167, "finished": true, "Ev Idx": 16067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960974.440, "dur": 1.330,
    "args": {
      "Python parent id": 12, "Python id": 13168, "finished": true, "Ev Idx": 16068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960974.636, "dur": 1.081,
    "args": {
      "Python parent id": 13168, "Python id": 13169, "finished": true, "Ev Idx": 16069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960976.823, "dur": 1.348,
    "args": {
      "Python parent id": 12, "Python id": 13170, "finished": true, "Ev Idx": 16070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960977.019, "dur": 1.098,
    "args": {
      "Python parent id": 13170, "Python id": 13171, "finished": true, "Ev Idx": 16071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960978.773, "dur": 1.326,
    "args": {
      "Python parent id": 12, "Python id": 13172, "finished": true, "Ev Idx": 16072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960978.967, "dur": 1.082,
    "args": {
      "Python parent id": 13172, "Python id": 13173, "finished": true, "Ev Idx": 16073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960980.712, "dur": 1.357,
    "args": {
      "Python parent id": 12, "Python id": 13174, "finished": true, "Ev Idx": 16074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960980.940, "dur": 1.080,
    "args": {
      "Python parent id": 13174, "Python id": 13175, "finished": true, "Ev Idx": 16075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960982.660, "dur": 1.139,
    "args": {
      "Python parent id": 12, "Python id": 13176, "finished": true, "Ev Idx": 16076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960982.855, "dur": 0.899,
    "args": {
      "Python parent id": 13176, "Python id": 13177, "finished": true, "Ev Idx": 16077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960984.454, "dur": 1.290,
    "args": {
      "Python parent id": 12, "Python id": 13178, "finished": true, "Ev Idx": 16078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960984.652, "dur": 1.043,
    "args": {
      "Python parent id": 13178, "Python id": 13179, "finished": true, "Ev Idx": 16079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960986.341, "dur": 1.204,
    "args": {
      "Python parent id": 12, "Python id": 13180, "finished": true, "Ev Idx": 16080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960986.571, "dur": 0.930,
    "args": {
      "Python parent id": 13180, "Python id": 13181, "finished": true, "Ev Idx": 16081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960988.178, "dur": 1.384,
    "args": {
      "Python parent id": 12, "Python id": 13182, "finished": true, "Ev Idx": 16082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960988.376, "dur": 1.136,
    "args": {
      "Python parent id": 13182, "Python id": 13183, "finished": true, "Ev Idx": 16083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960990.232, "dur": 1.300,
    "args": {
      "Python parent id": 12, "Python id": 13184, "finished": true, "Ev Idx": 16084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960990.428, "dur": 1.060,
    "args": {
      "Python parent id": 13184, "Python id": 13185, "finished": true, "Ev Idx": 16085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960992.111, "dur": 1.415,
    "args": {
      "Python parent id": 12, "Python id": 13186, "finished": true, "Ev Idx": 16086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960992.306, "dur": 1.175,
    "args": {
      "Python parent id": 13186, "Python id": 13187, "finished": true, "Ev Idx": 16087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960994.124, "dur": 1.448,
    "args": {
      "Python parent id": 12, "Python id": 13188, "finished": true, "Ev Idx": 16088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960994.319, "dur": 1.206,
    "args": {
      "Python parent id": 13188, "Python id": 13189, "finished": true, "Ev Idx": 16089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960996.174, "dur": 1.319,
    "args": {
      "Python parent id": 12, "Python id": 13190, "finished": true, "Ev Idx": 16090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960996.428, "dur": 1.012,
    "args": {
      "Python parent id": 13190, "Python id": 13191, "finished": true, "Ev Idx": 16091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740960998.132, "dur": 1.406,
    "args": {
      "Python parent id": 12, "Python id": 13192, "finished": true, "Ev Idx": 16092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740960998.327, "dur": 1.161,
    "args": {
      "Python parent id": 13192, "Python id": 13193, "finished": true, "Ev Idx": 16093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961000.597, "dur": 1.315,
    "args": {
      "Python parent id": 12, "Python id": 13194, "finished": true, "Ev Idx": 16094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961000.852, "dur": 1.011,
    "args": {
      "Python parent id": 13194, "Python id": 13195, "finished": true, "Ev Idx": 16095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961002.515, "dur": 1.418,
    "args": {
      "Python parent id": 12, "Python id": 13196, "finished": true, "Ev Idx": 16096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961002.708, "dur": 1.179,
    "args": {
      "Python parent id": 13196, "Python id": 13197, "finished": true, "Ev Idx": 16097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961004.547, "dur": 1.164,
    "args": {
      "Python parent id": 12, "Python id": 13198, "finished": true, "Ev Idx": 16098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961004.745, "dur": 0.914,
    "args": {
      "Python parent id": 13198, "Python id": 13199, "finished": true, "Ev Idx": 16099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961006.293, "dur": 1.324,
    "args": {
      "Python parent id": 12, "Python id": 13200, "finished": true, "Ev Idx": 16100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961006.488, "dur": 1.076,
    "args": {
      "Python parent id": 13200, "Python id": 13201, "finished": true, "Ev Idx": 16101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961008.253, "dur": 1.268,
    "args": {
      "Python parent id": 12, "Python id": 13202, "finished": true, "Ev Idx": 16102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961008.449, "dur": 1.022,
    "args": {
      "Python parent id": 13202, "Python id": 13203, "finished": true, "Ev Idx": 16103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961010.208, "dur": 1.307,
    "args": {
      "Python parent id": 12, "Python id": 13204, "finished": true, "Ev Idx": 16104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961010.402, "dur": 1.067,
    "args": {
      "Python parent id": 13204, "Python id": 13205, "finished": true, "Ev Idx": 16105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961012.225, "dur": 1.178,
    "args": {
      "Python parent id": 12, "Python id": 13206, "finished": true, "Ev Idx": 16106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961012.421, "dur": 0.929,
    "args": {
      "Python parent id": 13206, "Python id": 13207, "finished": true, "Ev Idx": 16107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961013.996, "dur": 1.271,
    "args": {
      "Python parent id": 12, "Python id": 13208, "finished": true, "Ev Idx": 16108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961014.191, "dur": 1.025,
    "args": {
      "Python parent id": 13208, "Python id": 13209, "finished": true, "Ev Idx": 16109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961015.896, "dur": 1.163,
    "args": {
      "Python parent id": 12, "Python id": 13210, "finished": true, "Ev Idx": 16110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961016.092, "dur": 0.921,
    "args": {
      "Python parent id": 13210, "Python id": 13211, "finished": true, "Ev Idx": 16111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961017.723, "dur": 1.265,
    "args": {
      "Python parent id": 12, "Python id": 13212, "finished": true, "Ev Idx": 16112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961017.917, "dur": 1.027,
    "args": {
      "Python parent id": 13212, "Python id": 13213, "finished": true, "Ev Idx": 16113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961019.728, "dur": 1.559,
    "args": {
      "Python parent id": 12, "Python id": 13214, "finished": true, "Ev Idx": 16114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961019.924, "dur": 1.313,
    "args": {
      "Python parent id": 13214, "Python id": 13215, "finished": true, "Ev Idx": 16115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961021.889, "dur": 1.381,
    "args": {
      "Python parent id": 12, "Python id": 13216, "finished": true, "Ev Idx": 16116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961022.125, "dur": 1.093,
    "args": {
      "Python parent id": 13216, "Python id": 13217, "finished": true, "Ev Idx": 16117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961025.714, "dur": 1.618,
    "args": {
      "Python parent id": 12, "Python id": 13218, "finished": true, "Ev Idx": 16118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961025.998, "dur": 1.283,
    "args": {
      "Python parent id": 13218, "Python id": 13219, "finished": true, "Ev Idx": 16119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961027.943, "dur": 1.311,
    "args": {
      "Python parent id": 12, "Python id": 13220, "finished": true, "Ev Idx": 16120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961028.138, "dur": 1.064,
    "args": {
      "Python parent id": 13220, "Python id": 13221, "finished": true, "Ev Idx": 16121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961029.840, "dur": 1.300,
    "args": {
      "Python parent id": 12, "Python id": 13222, "finished": true, "Ev Idx": 16122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961030.037, "dur": 1.054,
    "args": {
      "Python parent id": 13222, "Python id": 13223, "finished": true, "Ev Idx": 16123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961031.754, "dur": 1.318,
    "args": {
      "Python parent id": 12, "Python id": 13224, "finished": true, "Ev Idx": 16124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961031.948, "dur": 1.075,
    "args": {
      "Python parent id": 13224, "Python id": 13225, "finished": true, "Ev Idx": 16125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961033.672, "dur": 1.318,
    "args": {
      "Python parent id": 12, "Python id": 13226, "finished": true, "Ev Idx": 16126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961033.874, "dur": 1.067,
    "args": {
      "Python parent id": 13226, "Python id": 13227, "finished": true, "Ev Idx": 16127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961035.615, "dur": 1.514,
    "args": {
      "Python parent id": 12, "Python id": 13228, "finished": true, "Ev Idx": 16128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961035.810, "dur": 1.274,
    "args": {
      "Python parent id": 13228, "Python id": 13229, "finished": true, "Ev Idx": 16129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961037.711, "dur": 1.400,
    "args": {
      "Python parent id": 12, "Python id": 13230, "finished": true, "Ev Idx": 16130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961037.907, "dur": 1.154,
    "args": {
      "Python parent id": 13230, "Python id": 13231, "finished": true, "Ev Idx": 16131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961039.828, "dur": 1.380,
    "args": {
      "Python parent id": 12, "Python id": 13232, "finished": true, "Ev Idx": 16132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961040.025, "dur": 1.139,
    "args": {
      "Python parent id": 13232, "Python id": 13233, "finished": true, "Ev Idx": 16133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961041.842, "dur": 1.408,
    "args": {
      "Python parent id": 12, "Python id": 13234, "finished": true, "Ev Idx": 16134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961042.041, "dur": 1.161,
    "args": {
      "Python parent id": 13234, "Python id": 13235, "finished": true, "Ev Idx": 16135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961043.872, "dur": 1.413,
    "args": {
      "Python parent id": 12, "Python id": 13236, "finished": true, "Ev Idx": 16136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961044.068, "dur": 1.172,
    "args": {
      "Python parent id": 13236, "Python id": 13237, "finished": true, "Ev Idx": 16137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961045.915, "dur": 1.439,
    "args": {
      "Python parent id": 12, "Python id": 13238, "finished": true, "Ev Idx": 16138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961046.114, "dur": 1.190,
    "args": {
      "Python parent id": 13238, "Python id": 13239, "finished": true, "Ev Idx": 16139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961047.940, "dur": 1.356,
    "args": {
      "Python parent id": 12, "Python id": 13240, "finished": true, "Ev Idx": 16140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961048.135, "dur": 1.111,
    "args": {
      "Python parent id": 13240, "Python id": 13241, "finished": true, "Ev Idx": 16141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961050.166, "dur": 1.554,
    "args": {
      "Python parent id": 12, "Python id": 13242, "finished": true, "Ev Idx": 16142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961050.363, "dur": 1.305,
    "args": {
      "Python parent id": 13242, "Python id": 13243, "finished": true, "Ev Idx": 16143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961052.441, "dur": 1.334,
    "args": {
      "Python parent id": 12, "Python id": 13244, "finished": true, "Ev Idx": 16144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961052.675, "dur": 1.051,
    "args": {
      "Python parent id": 13244, "Python id": 13245, "finished": true, "Ev Idx": 16145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961054.357, "dur": 1.293,
    "args": {
      "Python parent id": 12, "Python id": 13246, "finished": true, "Ev Idx": 16146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961054.553, "dur": 1.045,
    "args": {
      "Python parent id": 13246, "Python id": 13247, "finished": true, "Ev Idx": 16147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961056.251, "dur": 1.319,
    "args": {
      "Python parent id": 12, "Python id": 13248, "finished": true, "Ev Idx": 16148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961056.446, "dur": 1.073,
    "args": {
      "Python parent id": 13248, "Python id": 13249, "finished": true, "Ev Idx": 16149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961058.301, "dur": 1.349,
    "args": {
      "Python parent id": 12, "Python id": 13250, "finished": true, "Ev Idx": 16150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961058.500, "dur": 1.100,
    "args": {
      "Python parent id": 13250, "Python id": 13251, "finished": true, "Ev Idx": 16151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961060.254, "dur": 1.521,
    "args": {
      "Python parent id": 12, "Python id": 13252, "finished": true, "Ev Idx": 16152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961060.452, "dur": 1.273,
    "args": {
      "Python parent id": 13252, "Python id": 13253, "finished": true, "Ev Idx": 16153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961062.362, "dur": 1.382,
    "args": {
      "Python parent id": 12, "Python id": 13254, "finished": true, "Ev Idx": 16154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961062.559, "dur": 1.140,
    "args": {
      "Python parent id": 13254, "Python id": 13255, "finished": true, "Ev Idx": 16155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961064.432, "dur": 1.465,
    "args": {
      "Python parent id": 12, "Python id": 13256, "finished": true, "Ev Idx": 16156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961064.626, "dur": 1.225,
    "args": {
      "Python parent id": 13256, "Python id": 13257, "finished": true, "Ev Idx": 16157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961066.479, "dur": 1.367,
    "args": {
      "Python parent id": 12, "Python id": 13258, "finished": true, "Ev Idx": 16158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961066.713, "dur": 1.088,
    "args": {
      "Python parent id": 13258, "Python id": 13259, "finished": true, "Ev Idx": 16159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961068.424, "dur": 1.373,
    "args": {
      "Python parent id": 12, "Python id": 13260, "finished": true, "Ev Idx": 16160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961068.621, "dur": 1.127,
    "args": {
      "Python parent id": 13260, "Python id": 13261, "finished": true, "Ev Idx": 16161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/streams.py(210): elapsed_time", "pid": 49625, "tid": 49625,
    "ts": 7253740961070.381, "dur": 1.338,
    "args": {
      "Python parent id": 12, "Python id": 13262, "finished": true, "Ev Idx": 16162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method elapsed_time of Event object at 0x7f6dc801a580>", "pid": 49625, "tid": 49625,
    "ts": 7253740961070.654, "dur": 1.017,
    "args": {
      "Python parent id": 13262, "Python id": 13263, "finished": true, "Ev Idx": 16163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(3778): _mean_dispatcher", "pid": 49625, "tid": 49625,
    "ts": 7253740961078.200, "dur": 0.158,
    "args": {
      "Python parent id": 12, "Python id": 13264, "finished": true, "Ev Idx": 16164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(3783): mean", "pid": 49625, "tid": 49625,
    "ts": 7253740961083.015, "dur": 135.522,
    "args": {
      "Python parent id": 12, "Python id": 13265, "finished": true, "Ev Idx": 16165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/_methods.py(119): _mean", "pid": 49625, "tid": 49625,
    "ts": 7253740961097.086, "dur": 107.689,
    "args": {
      "Python parent id": 13265, "Python id": 13266, "finished": true, "Ev Idx": 16166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function asanyarray>", "pid": 49625, "tid": 49625,
    "ts": 7253740961099.311, "dur": 23.649,
    "args": {
      "Python parent id": 13266, "Python id": 13267, "finished": true, "Ev Idx": 16167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/_methods.py(77): _count_reduce_items", "pid": 49625, "tid": 49625,
    "ts": 7253740961126.429, "dur": 19.280,
    "args": {
      "Python parent id": 13266, "Python id": 13268, "finished": true, "Ev Idx": 16168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function normalize_axis_index>", "pid": 49625, "tid": 49625,
    "ts": 7253740961132.902, "dur": 2.573,
    "args": {
      "Python parent id": 13268, "Python id": 13269, "finished": true, "Ev Idx": 16169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961154.163, "dur": 0.611,
    "args": {
      "Python parent id": 13266, "Python id": 13270, "finished": true, "Ev Idx": 16170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961156.586, "dur": 0.094,
    "args": {
      "Python parent id": 13266, "Python id": 13271, "finished": true, "Ev Idx": 16171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reduce of numpy.ufunc object at 0x7f789a9c0a40>", "pid": 49625, "tid": 49625,
    "ts": 7253740961160.204, "dur": 34.712,
    "args": {
      "Python parent id": 13266, "Python id": 13272, "finished": true, "Ev Idx": 16172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961196.027, "dur": 1.118,
    "args": {
      "Python parent id": 13266, "Python id": 13273, "finished": true, "Ev Idx": 16173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961198.337, "dur": 2.855,
    "args": {
      "Python parent id": 13266, "Python id": 13274, "finished": true, "Ev Idx": 16174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/lib/_function_base_impl.py(3865): _median_dispatcher", "pid": 49625, "tid": 49625,
    "ts": 7253740961223.172, "dur": 0.438,
    "args": {
      "Python parent id": 12, "Python id": 13275, "finished": true, "Ev Idx": 16175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/lib/_function_base_impl.py(3870): median", "pid": 49625, "tid": 49625,
    "ts": 7253740961225.750, "dur": 285.824,
    "args": {
      "Python parent id": 12, "Python id": 13276, "finished": true, "Ev Idx": 16176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/lib/_function_base_impl.py(3788): _ureduce", "pid": 49625, "tid": 49625,
    "ts": 7253740961232.389, "dur": 278.670,
    "args": {
      "Python parent id": 13276, "Python id": 13277, "finished": true, "Ev Idx": 16177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function asanyarray>", "pid": 49625, "tid": 49625,
    "ts": 7253740961234.621, "dur": 4.527,
    "args": {
      "Python parent id": 13277, "Python id": 13278, "finished": true, "Ev Idx": 16178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method get of dict object at 0x7f6dadbb7800>", "pid": 49625, "tid": 49625,
    "ts": 7253740961241.911, "dur": 0.360,
    "args": {
      "Python parent id": 13277, "Python id": 13279, "finished": true, "Ev Idx": 16179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method get of dict object at 0x7f6dadbb7800>", "pid": 49625, "tid": 49625,
    "ts": 7253740961243.075, "dur": 0.172,
    "args": {
      "Python parent id": 13277, "Python id": 13280, "finished": true, "Ev Idx": 16180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/lib/_function_base_impl.py(3965): _median", "pid": 49625, "tid": 49625,
    "ts": 7253740961249.864, "dur": 260.554,
    "args": {
      "Python parent id": 13277, "Python id": 13281, "finished": true, "Ev Idx": 16181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function asanyarray>", "pid": 49625, "tid": 49625,
    "ts": 7253740961251.434, "dur": 0.107,
    "args": {
      "Python parent id": 13281, "Python id": 13282, "finished": true, "Ev Idx": 16182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/numerictypes.py(471): issubdtype", "pid": 49625, "tid": 49625,
    "ts": 7253740961257.224, "dur": 8.674,
    "args": {
      "Python parent id": 13281, "Python id": 13283, "finished": true, "Ev Idx": 16183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/numerictypes.py(289): issubclass_", "pid": 49625, "tid": 49625,
    "ts": 7253740961259.126, "dur": 3.037,
    "args": {
      "Python parent id": 13283, "Python id": 13284, "finished": true, "Ev Idx": 16184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961259.932, "dur": 1.406,
    "args": {
      "Python parent id": 13284, "Python id": 13285, "finished": true, "Ev Idx": 16185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/numerictypes.py(289): issubclass_", "pid": 49625, "tid": 49625,
    "ts": 7253740961264.401, "dur": 0.700,
    "args": {
      "Python parent id": 13283, "Python id": 13286, "finished": true, "Ev Idx": 16186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961264.636, "dur": 0.355,
    "args": {
      "Python parent id": 13286, "Python id": 13287, "finished": true, "Ev Idx": 16187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961265.756, "dur": 0.085,
    "args": {
      "Python parent id": 13283, "Python id": 13288, "finished": true, "Ev Idx": 16188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7f6dc820e2c0>", "pid": 49625, "tid": 49625,
    "ts": 7253740961350.438, "dur": 0.385,
    "args": {
      "Python parent id": 13281, "Python id": 13289, "finished": true, "Ev Idx": 16189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(754): _partition_dispatcher", "pid": 49625, "tid": 49625,
    "ts": 7253740961353.883, "dur": 0.133,
    "args": {
      "Python parent id": 13281, "Python id": 13290, "finished": true, "Ev Idx": 16190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(758): partition", "pid": 49625, "tid": 49625,
    "ts": 7253740961356.055, "dur": 68.977,
    "args": {
      "Python parent id": 13281, "Python id": 13291, "finished": true, "Ev Idx": 16191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function asanyarray>", "pid": 49625, "tid": 49625,
    "ts": 7253740961356.795, "dur": 0.070,
    "args": {
      "Python parent id": 13291, "Python id": 13292, "finished": true, "Ev Idx": 16192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method flatten of numpy.ndarray object at 0x7f6dadb3da10>", "pid": 49625, "tid": 49625,
    "ts": 7253740961359.057, "dur": 6.064,
    "args": {
      "Python parent id": 13291, "Python id": 13293, "finished": true, "Ev Idx": 16193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method partition of numpy.ndarray object at 0x7f6dadba52f0>", "pid": 49625, "tid": 49625,
    "ts": 7253740961367.355, "dur": 57.499,
    "args": {
      "Python parent id": 13291, "Python id": 13294, "finished": true, "Ev Idx": 16194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(3778): _mean_dispatcher", "pid": 49625, "tid": 49625,
    "ts": 7253740961434.231, "dur": 0.157,
    "args": {
      "Python parent id": 13281, "Python id": 13295, "finished": true, "Ev Idx": 16195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/fromnumeric.py(3783): mean", "pid": 49625, "tid": 49625,
    "ts": 7253740961435.618, "dur": 21.159,
    "args": {
      "Python parent id": 13281, "Python id": 13296, "finished": true, "Ev Idx": 16196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/_methods.py(119): _mean", "pid": 49625, "tid": 49625,
    "ts": 7253740961438.780, "dur": 17.671,
    "args": {
      "Python parent id": 13296, "Python id": 13297, "finished": true, "Ev Idx": 16197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function asanyarray>", "pid": 49625, "tid": 49625,
    "ts": 7253740961439.067, "dur": 0.114,
    "args": {
      "Python parent id": 13297, "Python id": 13298, "finished": true, "Ev Idx": 16198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/_methods.py(77): _count_reduce_items", "pid": 49625, "tid": 49625,
    "ts": 7253740961439.788, "dur": 4.587,
    "args": {
      "Python parent id": 13297, "Python id": 13299, "finished": true, "Ev Idx": 16199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961440.849, "dur": 0.195,
    "args": {
      "Python parent id": 13299, "Python id": 13300, "finished": true, "Ev Idx": 16200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function normalize_axis_index>", "pid": 49625, "tid": 49625,
    "ts": 7253740961442.226, "dur": 0.330,
    "args": {
      "Python parent id": 13299, "Python id": 13301, "finished": true, "Ev Idx": 16201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961446.128, "dur": 0.176,
    "args": {
      "Python parent id": 13297, "Python id": 13302, "finished": true, "Ev Idx": 16202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function issubclass>", "pid": 49625, "tid": 49625,
    "ts": 7253740961447.051, "dur": 0.152,
    "args": {
      "Python parent id": 13297, "Python id": 13303, "finished": true, "Ev Idx": 16203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reduce of numpy.ufunc object at 0x7f789a9c0a40>", "pid": 49625, "tid": 49625,
    "ts": 7253740961447.729, "dur": 5.571,
    "args": {
      "Python parent id": 13297, "Python id": 13304, "finished": true, "Ev Idx": 16204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961453.886, "dur": 0.252,
    "args": {
      "Python parent id": 13297, "Python id": 13305, "finished": true, "Ev Idx": 16205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961454.743, "dur": 0.380,
    "args": {
      "Python parent id": 13297, "Python id": 13306, "finished": true, "Ev Idx": 16206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/lib/_utils_impl.py(644): _median_nancheck", "pid": 49625, "tid": 49625,
    "ts": 7253740961463.257, "dur": 46.932,
    "args": {
      "Python parent id": 13281, "Python id": 13307, "finished": true, "Ev Idx": 16207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method take of numpy.ndarray object at 0x7f6dadba52f0>", "pid": 49625, "tid": 49625,
    "ts": 7253740961466.076, "dur": 3.563,
    "args": {
      "Python parent id": 13307, "Python id": 13308, "finished": true, "Ev Idx": 16208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/ma/core.py(6670): isMaskedArray", "pid": 49625, "tid": 49625,
    "ts": 7253740961482.055, "dur": 2.395,
    "args": {
      "Python parent id": 13307, "Python id": 13309, "finished": true, "Ev Idx": 16209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961483.469, "dur": 0.872,
    "args": {
      "Python parent id": 13309, "Python id": 13310, "finished": true, "Ev Idx": 16210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method any of numpy.bool object at 0x7f77521c57c0>", "pid": 49625, "tid": 49625,
    "ts": 7253740961487.504, "dur": 22.496,
    "args": {
      "Python parent id": 13307, "Python id": 13311, "finished": true, "Ev Idx": 16211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "numpy/_core/_methods.py(59): _any", "pid": 49625, "tid": 49625,
    "ts": 7253740961496.462, "dur": 13.139,
    "args": {
      "Python parent id": 13311, "Python id": 13312, "finished": true, "Ev Idx": 16212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reduce of numpy.ufunc object at 0x7f789a9c0a40>", "pid": 49625, "tid": 49625,
    "ts": 7253740961499.403, "dur": 10.053,
    "args": {
      "Python parent id": 13312, "Python id": 13313, "finished": true, "Ev Idx": 16213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(803): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740961514.438, "dur": 70.013,
    "args": {
      "Python parent id": 12, "Python id": 13314, "finished": true, "Ev Idx": 16214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(817): stop", "pid": 49625, "tid": 49625,
    "ts": 7253740961516.507, "dur": 67.944,
    "args": {
      "Python parent id": 13314, "Python id": 13315, "finished": true, "Ev Idx": 16215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(862): _transit_action", "pid": 49625, "tid": 49625,
    "ts": 7253740961518.360, "dur": 66.091,
    "args": {
      "Python parent id": 13315, "Python id": 13316, "finished": true, "Ev Idx": 16216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method get of dict object at 0x7f6dadbb7800>", "pid": 49625, "tid": 49625,
    "ts": 7253740961519.530, "dur": 4.242,
    "args": {
      "Python parent id": 13316, "Python id": 13317, "finished": true, "Ev Idx": 16217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1261): __hash__", "pid": 49625, "tid": 49625,
    "ts": 7253740961521.595, "dur": 1.592,
    "args": {
      "Python parent id": 13317, "Python id": 13318, "finished": true, "Ev Idx": 16218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hash>", "pid": 49625, "tid": 49625,
    "ts": 7253740961522.951, "dur": 0.148,
    "args": {
      "Python parent id": 13318, "Python id": 13319, "finished": true, "Ev Idx": 16219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(239): stop_trace", "pid": 49625, "tid": 49625,
    "ts": 7253740961525.806, "dur": 58.645,
    "args": {
      "Python parent id": 13316, "Python id": 13320, "finished": true, "Ev Idx": 16220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/profiler.py(375): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740961530.105, "dur": 54.346,
    "args": {
      "Python parent id": 13320, "Python id": 13321, "finished": true, "Ev Idx": 16221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961531.668, "dur": 0.322,
    "args": {
      "Python parent id": 13321, "Python id": 13322, "finished": true, "Ev Idx": 16222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961533.226, "dur": 0.182,
    "args": {
      "Python parent id": 13321, "Python id": 13323, "finished": true, "Ev Idx": 16223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961534.084, "dur": 0.089,
    "args": {
      "Python parent id": 13321, "Python id": 13324, "finished": true, "Ev Idx": 16224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1030): synchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740961535.370, "dur": 49.081,
    "args": {
      "Python parent id": 13321, "Python id": 13325, "finished": true, "Ev Idx": 16225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740961535.866, "dur": 1.462,
    "args": {
      "Python parent id": 13325, "Python id": 13326, "finished": true, "Ev Idx": 16226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740961536.304, "dur": 0.846,
    "args": {
      "Python parent id": 13326, "Python id": 13327, "finished": true, "Ev Idx": 16227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740961536.610, "dur": 0.434,
    "args": {
      "Python parent id": 13327, "Python id": 13328, "finished": true, "Ev Idx": 16228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(490): __init__", "pid": 49625, "tid": 49625,
    "ts": 7253740961538.438, "dur": 22.134,
    "args": {
      "Python parent id": 13325, "Python id": 13329, "finished": true, "Ev Idx": 16229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740961538.908, "dur": 21.441,
    "args": {
      "Python parent id": 13329, "Python id": 13330, "finished": true, "Ev Idx": 16230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961539.429, "dur": 0.179,
    "args": {
      "Python parent id": 13330, "Python id": 13331, "finished": true, "Ev Idx": 16231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961539.863, "dur": 0.049,
    "args": {
      "Python parent id": 13330, "Python id": 13332, "finished": true, "Ev Idx": 16232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961540.402, "dur": 0.052,
    "args": {
      "Python parent id": 13330, "Python id": 13333, "finished": true, "Ev Idx": 16233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740961541.602, "dur": 0.090,
    "args": {
      "Python parent id": 13330, "Python id": 13334, "finished": true, "Ev Idx": 16234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961542.191, "dur": 0.163,
    "args": {
      "Python parent id": 13330, "Python id": 13335, "finished": true, "Ev Idx": 16235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740961543.115, "dur": 17.093,
    "args": {
      "Python parent id": 13330, "Python id": 13336, "finished": true, "Ev Idx": 16236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961543.439, "dur": 0.069,
    "args": {
      "Python parent id": 13336, "Python id": 13337, "finished": true, "Ev Idx": 16237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961543.833, "dur": 0.078,
    "args": {
      "Python parent id": 13336, "Python id": 13338, "finished": true, "Ev Idx": 16238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961544.381, "dur": 0.072,
    "args": {
      "Python parent id": 13336, "Python id": 13339, "finished": true, "Ev Idx": 16239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 49625, "tid": 49625,
    "ts": 7253740961545.079, "dur": 0.032,
    "args": {
      "Python parent id": 13336, "Python id": 13340, "finished": true, "Ev Idx": 16240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 49625, "tid": 49625,
    "ts": 7253740961545.798, "dur": 14.220,
    "args": {
      "Python parent id": 13336, "Python id": 13341, "finished": true, "Ev Idx": 16241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 49625, "tid": 49625,
    "ts": 7253740961546.514, "dur": 13.307,
    "args": {
      "Python parent id": 13341, "Python id": 13342, "finished": true, "Ev Idx": 16242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 49625, "tid": 49625,
    "ts": 7253740961546.818, "dur": 8.604,
    "args": {
      "Python parent id": 13342, "Python id": 13343, "finished": true, "Ev Idx": 16243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 49625, "tid": 49625,
    "ts": 7253740961547.216, "dur": 8.140,
    "args": {
      "Python parent id": 13343, "Python id": 13344, "finished": true, "Ev Idx": 16244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 49625, "tid": 49625,
    "ts": 7253740961547.504, "dur": 0.549,
    "args": {
      "Python parent id": 13344, "Python id": 13345, "finished": true, "Ev Idx": 16245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 49625, "tid": 49625,
    "ts": 7253740961547.808, "dur": 0.070,
    "args": {
      "Python parent id": 13345, "Python id": 13346, "finished": true, "Ev Idx": 16246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 49625, "tid": 49625,
    "ts": 7253740961548.380, "dur": 6.183,
    "args": {
      "Python parent id": 13344, "Python id": 13347, "finished": true, "Ev Idx": 16247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 49625, "tid": 49625,
    "ts": 7253740961548.996, "dur": 5.304,
    "args": {
      "Python parent id": 13347, "Python id": 13348, "finished": true, "Ev Idx": 16248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 49625, "tid": 49625,
    "ts": 7253740961549.722, "dur": 4.456,
    "args": {
      "Python parent id": 13348, "Python id": 13349, "finished": true, "Ev Idx": 16249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 49625, "tid": 49625,
    "ts": 7253740961550.118, "dur": 3.204,
    "args": {
      "Python parent id": 13349, "Python id": 13350, "finished": true, "Ev Idx": 16250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 49625, "tid": 49625,
    "ts": 7253740961550.712, "dur": 1.196,
    "args": {
      "Python parent id": 13350, "Python id": 13351, "finished": true, "Ev Idx": 16251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 49625, "tid": 49625,
    "ts": 7253740961550.913, "dur": 0.063,
    "args": {
      "Python parent id": 13351, "Python id": 13352, "finished": true, "Ev Idx": 16252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7f781bf90620>", "pid": 49625, "tid": 49625,
    "ts": 7253740961551.414, "dur": 0.406,
    "args": {
      "Python parent id": 13351, "Python id": 13353, "finished": true, "Ev Idx": 16253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 49625, "tid": 49625,
    "ts": 7253740961554.968, "dur": 0.316,
    "args": {
      "Python parent id": 13344, "Python id": 13354, "finished": true, "Ev Idx": 16254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7f78ea906e20>", "pid": 49625, "tid": 49625,
    "ts": 7253740961555.848, "dur": 0.224,
    "args": {
      "Python parent id": 13342, "Python id": 13355, "finished": true, "Ev Idx": 16255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 49625, "tid": 49625,
    "ts": 7253740961556.754, "dur": 2.984,
    "args": {
      "Python parent id": 13342, "Python id": 13356, "finished": true, "Ev Idx": 16256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 49625, "tid": 49625,
    "ts": 7253740961557.138, "dur": 2.540,
    "args": {
      "Python parent id": 13356, "Python id": 13357, "finished": true, "Ev Idx": 16257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 49625, "tid": 49625,
    "ts": 7253740961557.536, "dur": 0.547,
    "args": {
      "Python parent id": 13357, "Python id": 13358, "finished": true, "Ev Idx": 16258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 49625, "tid": 49625,
    "ts": 7253740961557.718, "dur": 0.277,
    "args": {
      "Python parent id": 13358, "Python id": 13359, "finished": true, "Ev Idx": 16259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 49625, "tid": 49625,
    "ts": 7253740961557.882, "dur": 0.080,
    "args": {
      "Python parent id": 13359, "Python id": 13360, "finished": true, "Ev Idx": 16260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740961558.336, "dur": 1.283,
    "args": {
      "Python parent id": 13357, "Python id": 13361, "finished": true, "Ev Idx": 16261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(494): __enter__", "pid": 49625, "tid": 49625,
    "ts": 7253740961561.321, "dur": 1.267,
    "args": {
      "Python parent id": 13325, "Python id": 13362, "finished": true, "Ev Idx": 16262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_exchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740961561.730, "dur": 0.739,
    "args": {
      "Python parent id": 13362, "Python id": 13363, "finished": true, "Ev Idx": 16263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_synchronize>", "pid": 49625, "tid": 49625,
    "ts": 7253740961563.069, "dur": 5.431,
    "args": {
      "Python parent id": 13325, "Python id": 13364, "finished": true, "Ev Idx": 16264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(497): __exit__", "pid": 49625, "tid": 49625,
    "ts": 7253740961569.081, "dur": 15.370,
    "args": {
      "Python parent id": 13325, "Python id": 13365, "finished": true, "Ev Idx": 16265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_maybeExchangeDevice>", "pid": 49625, "tid": 49625,
    "ts": 7253740961569.361, "dur": 0.374,
    "args": {
      "Python parent id": 13365, "Python id": 13366, "finished": true, "Ev Idx": 16266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function perf_counter_ns>", "pid": 49625, "tid": 49625,
    "ts": 7253740961572.175, "dur": 0.704,
    "args": {
      "Python parent id": 13365, "Python id": 13367, "finished": true, "Ev Idx": 16267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _disable_profiler of PyCapsule object at 0x7f76b136f990>", "pid": 49625, "tid": 49625,
    "ts": 7253740961575.824, "dur": 8.627,
    "args": {
      "Python parent id": 13365, "Python id": 13368, "finished": true, "Ev Idx": 16268
    }
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 49625, "tid": 49625,
    "ts": 7253740836911.110, "dur": 2.038,
    "args": {
      "External id": 2,      "cid": 290, "correlation": 1
    }
  },
  {
    "ph": "f", "id": 1, "pid": 49625, "tid": 49625, "ts": 7253740836911.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 49625, "tid": 49625,
    "ts": 7253740836915.833, "dur": 117.362,
    "args": {
      "External id": 2,            "bytes": 134217728, "cid": 108, "correlation": 2, "ptr": "0x55916b31e260"
    }
  },
  {
    "ph": "s", "id": 2, "pid": 49625, "tid": 49625, "ts": 7253740836915.833,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740837058.228, "dur": 16.320,
    "args": {
      "External id": 3,      "cid": 134, "correlation": 3, "src": "0x7f6dd4e00000", "dst": "0x7f6d67e00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 3, "pid": 49625, "tid": 49625, "ts": 7253740837058.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740837249.901, "dur": 7746.938,
    "args": {
      "External id": 10,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 4,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 4, "pid": 49625, "tid": 49625, "ts": 7253740837249.901,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740845077.105, "dur": 5696.986,
    "args": {
      "External id": 11,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 5,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 5, "pid": 49625, "tid": 49625, "ts": 7253740845077.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 49625, "tid": 49625,
    "ts": 7253740850833.856, "dur": 0.754,
    "args": {
      "External id": 12,      "cid": 290, "correlation": 6
    }
  },
  {
    "ph": "f", "id": 6, "pid": 49625, "tid": 49625, "ts": 7253740850833.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 49625, "tid": 49625,
    "ts": 7253740850835.921, "dur": 60.288,
    "args": {
      "External id": 12,            "bytes": 134217728, "cid": 108, "correlation": 7, "ptr": "0x55916bd67bc0"
    }
  },
  {
    "ph": "s", "id": 7, "pid": 49625, "tid": 49625, "ts": 7253740850835.921,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740850906.367, "dur": 7.890,
    "args": {
      "External id": 12,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 8,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 8, "pid": 49625, "tid": 49625, "ts": 7253740850906.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 49625, "tid": 49625,
    "ts": 7253740850945.330, "dur": 0.307,
    "args": {
      "External id": 13,      "cid": 290, "correlation": 9
    }
  },
  {
    "ph": "f", "id": 9, "pid": 49625, "tid": 49625, "ts": 7253740850945.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 49625, "tid": 49625,
    "ts": 7253740850946.107, "dur": 31.263,
    "args": {
      "External id": 13,            "bytes": 134217728, "cid": 108, "correlation": 10, "ptr": "0x3"
    }
  },
  {
    "ph": "s", "id": 10, "pid": 49625, "tid": 49625, "ts": 7253740850946.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740850981.495, "dur": 7.503,
    "args": {
      "External id": 13,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 11,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 11, "pid": 49625, "tid": 49625, "ts": 7253740850981.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740851027.469, "dur": 10099.621,
    "args": {
      "External id": 14,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 12,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 12, "pid": 49625, "tid": 49625, "ts": 7253740851027.469,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740861230.260, "dur": 13590.295,
    "args": {
      "External id": 15,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 13,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 13, "pid": 49625, "tid": 49625, "ts": 7253740861230.260,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740874884.253, "dur": 6.941,
    "args": {
      "External id": 17,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 14,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 14, "pid": 49625, "tid": 49625, "ts": 7253740874884.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 49625, "tid": 49625,
    "ts": 7253740874901.761, "dur": 0.564,
    "args": {
      "External id": 18,      "cid": 290, "correlation": 15
    }
  },
  {
    "ph": "f", "id": 15, "pid": 49625, "tid": 49625, "ts": 7253740874901.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 49625, "tid": 49625,
    "ts": 7253740874903.274, "dur": 60.209,
    "args": {
      "External id": 18,            "bytes": 134217728, "cid": 108, "correlation": 16, "ptr": "0xffffffff"
    }
  },
  {
    "ph": "s", "id": 16, "pid": 49625, "tid": 49625, "ts": 7253740874903.274,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740874967.551, "dur": 6.361,
    "args": {
      "External id": 18,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 17,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 17, "pid": 49625, "tid": 49625, "ts": 7253740874967.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875032.460, "dur": 9.224,
    "args": {
      "External id": 21,      "cid": 134, "correlation": 18, "src": "0x7f6d67e00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 18, "pid": 49625, "tid": 49625, "ts": 7253740875032.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875148.880, "dur": 6.630,
    "args": {
      "External id": 28,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 19,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 19, "pid": 49625, "tid": 49625, "ts": 7253740875148.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875168.249, "dur": 4.594,
    "args": {
      "External id": 29,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 20,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 20, "pid": 49625, "tid": 49625, "ts": 7253740875168.249,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875180.866, "dur": 3.513,
    "args": {
      "External id": 30,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 21,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 21, "pid": 49625, "tid": 49625, "ts": 7253740875180.866,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875192.492, "dur": 5.178,
    "args": {
      "External id": 31,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 22,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 22, "pid": 49625, "tid": 49625, "ts": 7253740875192.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875206.879, "dur": 4.047,
    "args": {
      "External id": 32,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 23,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 23, "pid": 49625, "tid": 49625, "ts": 7253740875206.879,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875222.686, "dur": 4.573,
    "args": {
      "External id": 33,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 24,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 24, "pid": 49625, "tid": 49625, "ts": 7253740875222.686,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875233.108, "dur": 4.329,
    "args": {
      "External id": 35,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 25,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 25, "pid": 49625, "tid": 49625, "ts": 7253740875233.108,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875241.825, "dur": 3.350,
    "args": {
      "External id": 36,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 26,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 26, "pid": 49625, "tid": 49625, "ts": 7253740875241.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875262.689, "dur": 5.388,
    "args": {
      "External id": 39,      "cid": 134, "correlation": 27, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 27, "pid": 49625, "tid": 49625, "ts": 7253740875262.689,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875304.543, "dur": 3.700,
    "args": {
      "External id": 46,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 28,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 28, "pid": 49625, "tid": 49625, "ts": 7253740875304.543,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875314.337, "dur": 3.713,
    "args": {
      "External id": 47,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 29,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 29, "pid": 49625, "tid": 49625, "ts": 7253740875314.337,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875323.153, "dur": 3.333,
    "args": {
      "External id": 48,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 30,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 30, "pid": 49625, "tid": 49625, "ts": 7253740875323.153,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875330.378, "dur": 3.982,
    "args": {
      "External id": 49,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 31,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 31, "pid": 49625, "tid": 49625, "ts": 7253740875330.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875339.297, "dur": 3.730,
    "args": {
      "External id": 50,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 32,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 32, "pid": 49625, "tid": 49625, "ts": 7253740875339.297,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875348.894, "dur": 3.075,
    "args": {
      "External id": 51,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 33,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 33, "pid": 49625, "tid": 49625, "ts": 7253740875348.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875356.686, "dur": 2.991,
    "args": {
      "External id": 53,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 34,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 34, "pid": 49625, "tid": 49625, "ts": 7253740875356.686,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875363.044, "dur": 3.466,
    "args": {
      "External id": 54,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 35,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 35, "pid": 49625, "tid": 49625, "ts": 7253740875363.044,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875379.484, "dur": 3.808,
    "args": {
      "External id": 57,      "cid": 134, "correlation": 36, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 36, "pid": 49625, "tid": 49625, "ts": 7253740875379.484,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875412.378, "dur": 3.614,
    "args": {
      "External id": 64,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 37,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 37, "pid": 49625, "tid": 49625, "ts": 7253740875412.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875420.972, "dur": 2.910,
    "args": {
      "External id": 65,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 38,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 38, "pid": 49625, "tid": 49625, "ts": 7253740875420.972,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875428.547, "dur": 2.793,
    "args": {
      "External id": 66,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 39,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 39, "pid": 49625, "tid": 49625, "ts": 7253740875428.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875435.076, "dur": 3.274,
    "args": {
      "External id": 67,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 40,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 40, "pid": 49625, "tid": 49625, "ts": 7253740875435.076,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875442.650, "dur": 3.258,
    "args": {
      "External id": 68,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 41,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 41, "pid": 49625, "tid": 49625, "ts": 7253740875442.650,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875451.078, "dur": 3.768,
    "args": {
      "External id": 69,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 42,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 42, "pid": 49625, "tid": 49625, "ts": 7253740875451.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875459.174, "dur": 3.121,
    "args": {
      "External id": 71,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 43,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 43, "pid": 49625, "tid": 49625, "ts": 7253740875459.174,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875465.565, "dur": 2.735,
    "args": {
      "External id": 72,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 44,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 44, "pid": 49625, "tid": 49625, "ts": 7253740875465.565,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875479.597, "dur": 3.495,
    "args": {
      "External id": 75,      "cid": 134, "correlation": 45, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 45, "pid": 49625, "tid": 49625, "ts": 7253740875479.597,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875510.996, "dur": 3.093,
    "args": {
      "External id": 82,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 46,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 46, "pid": 49625, "tid": 49625, "ts": 7253740875510.996,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875518.713, "dur": 3.028,
    "args": {
      "External id": 83,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 47,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 47, "pid": 49625, "tid": 49625, "ts": 7253740875518.713,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875526.342, "dur": 3.502,
    "args": {
      "External id": 84,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 48,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 48, "pid": 49625, "tid": 49625, "ts": 7253740875526.342,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875533.546, "dur": 3.388,
    "args": {
      "External id": 85,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 49,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 49, "pid": 49625, "tid": 49625, "ts": 7253740875533.546,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875541.035, "dur": 2.985,
    "args": {
      "External id": 86,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 50,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 50, "pid": 49625, "tid": 49625, "ts": 7253740875541.035,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875548.994, "dur": 3.207,
    "args": {
      "External id": 87,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 51,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 51, "pid": 49625, "tid": 49625, "ts": 7253740875548.994,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875556.273, "dur": 2.999,
    "args": {
      "External id": 89,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 52,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 52, "pid": 49625, "tid": 49625, "ts": 7253740875556.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875562.825, "dur": 2.955,
    "args": {
      "External id": 90,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 53,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 53, "pid": 49625, "tid": 49625, "ts": 7253740875562.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875576.683, "dur": 3.565,
    "args": {
      "External id": 93,      "cid": 134, "correlation": 54, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 54, "pid": 49625, "tid": 49625, "ts": 7253740875576.683,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875606.168, "dur": 3.508,
    "args": {
      "External id": 100,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 55,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 55, "pid": 49625, "tid": 49625, "ts": 7253740875606.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875614.101, "dur": 3.037,
    "args": {
      "External id": 101,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 56,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 56, "pid": 49625, "tid": 49625, "ts": 7253740875614.101,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875621.312, "dur": 3.224,
    "args": {
      "External id": 102,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 57,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 57, "pid": 49625, "tid": 49625, "ts": 7253740875621.312,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875628.009, "dur": 3.169,
    "args": {
      "External id": 103,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 58,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 58, "pid": 49625, "tid": 49625, "ts": 7253740875628.009,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875635.326, "dur": 3.228,
    "args": {
      "External id": 104,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 59,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 59, "pid": 49625, "tid": 49625, "ts": 7253740875635.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875643.444, "dur": 3.588,
    "args": {
      "External id": 105,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 60,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 60, "pid": 49625, "tid": 49625, "ts": 7253740875643.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875651.139, "dur": 3.262,
    "args": {
      "External id": 107,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 61,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 61, "pid": 49625, "tid": 49625, "ts": 7253740875651.139,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875657.614, "dur": 3.263,
    "args": {
      "External id": 108,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 62,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 62, "pid": 49625, "tid": 49625, "ts": 7253740875657.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875670.928, "dur": 3.436,
    "args": {
      "External id": 111,      "cid": 134, "correlation": 63, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 63, "pid": 49625, "tid": 49625, "ts": 7253740875670.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875699.846, "dur": 3.428,
    "args": {
      "External id": 118,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 64,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 64, "pid": 49625, "tid": 49625, "ts": 7253740875699.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875707.628, "dur": 3.746,
    "args": {
      "External id": 119,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 65,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 65, "pid": 49625, "tid": 49625, "ts": 7253740875707.628,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875729.854, "dur": 3.271,
    "args": {
      "External id": 120,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 66,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 66, "pid": 49625, "tid": 49625, "ts": 7253740875729.854,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875736.467, "dur": 3.458,
    "args": {
      "External id": 121,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 67,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 67, "pid": 49625, "tid": 49625, "ts": 7253740875736.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875744.042, "dur": 3.525,
    "args": {
      "External id": 122,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 68,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 68, "pid": 49625, "tid": 49625, "ts": 7253740875744.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875752.567, "dur": 3.265,
    "args": {
      "External id": 123,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 69,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 69, "pid": 49625, "tid": 49625, "ts": 7253740875752.567,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875759.837, "dur": 3.022,
    "args": {
      "External id": 125,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 70,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 70, "pid": 49625, "tid": 49625, "ts": 7253740875759.837,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875766.376, "dur": 3.584,
    "args": {
      "External id": 126,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 71,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 71, "pid": 49625, "tid": 49625, "ts": 7253740875766.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875779.908, "dur": 24.946,
    "args": {
      "External id": 129,      "cid": 134, "correlation": 72, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 72, "pid": 49625, "tid": 49625, "ts": 7253740875779.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875831.001, "dur": 3.725,
    "args": {
      "External id": 136,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 73,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 73, "pid": 49625, "tid": 49625, "ts": 7253740875831.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875839.535, "dur": 3.339,
    "args": {
      "External id": 137,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 74,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 74, "pid": 49625, "tid": 49625, "ts": 7253740875839.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875847.232, "dur": 3.338,
    "args": {
      "External id": 138,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 75,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 75, "pid": 49625, "tid": 49625, "ts": 7253740875847.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875854.059, "dur": 3.277,
    "args": {
      "External id": 139,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 76,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 76, "pid": 49625, "tid": 49625, "ts": 7253740875854.059,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875861.422, "dur": 3.347,
    "args": {
      "External id": 140,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 77,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 77, "pid": 49625, "tid": 49625, "ts": 7253740875861.422,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875869.612, "dur": 3.492,
    "args": {
      "External id": 141,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 78,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 78, "pid": 49625, "tid": 49625, "ts": 7253740875869.612,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875877.228, "dur": 3.595,
    "args": {
      "External id": 143,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 79,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 79, "pid": 49625, "tid": 49625, "ts": 7253740875877.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875884.107, "dur": 3.364,
    "args": {
      "External id": 144,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 80,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 80, "pid": 49625, "tid": 49625, "ts": 7253740875884.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875897.945, "dur": 3.634,
    "args": {
      "External id": 147,      "cid": 134, "correlation": 81, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 81, "pid": 49625, "tid": 49625, "ts": 7253740875897.945,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875926.304, "dur": 3.884,
    "args": {
      "External id": 154,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 82,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 82, "pid": 49625, "tid": 49625, "ts": 7253740875926.304,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875934.787, "dur": 3.603,
    "args": {
      "External id": 155,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 83,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 83, "pid": 49625, "tid": 49625, "ts": 7253740875934.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875942.634, "dur": 3.196,
    "args": {
      "External id": 156,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 84,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 84, "pid": 49625, "tid": 49625, "ts": 7253740875942.634,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875949.306, "dur": 4.270,
    "args": {
      "External id": 157,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 85,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 85, "pid": 49625, "tid": 49625, "ts": 7253740875949.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875958.037, "dur": 3.549,
    "args": {
      "External id": 158,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 86,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 86, "pid": 49625, "tid": 49625, "ts": 7253740875958.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875966.618, "dur": 3.068,
    "args": {
      "External id": 159,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 87,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 87, "pid": 49625, "tid": 49625, "ts": 7253740875966.618,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875973.825, "dur": 2.917,
    "args": {
      "External id": 161,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 88,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 88, "pid": 49625, "tid": 49625, "ts": 7253740875973.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740875979.978, "dur": 3.187,
    "args": {
      "External id": 162,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 89,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 89, "pid": 49625, "tid": 49625, "ts": 7253740875979.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740875993.184, "dur": 3.566,
    "args": {
      "External id": 165,      "cid": 134, "correlation": 90, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 90, "pid": 49625, "tid": 49625, "ts": 7253740875993.184,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876020.891, "dur": 4.050,
    "args": {
      "External id": 172,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 91,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 91, "pid": 49625, "tid": 49625, "ts": 7253740876020.891,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876029.624, "dur": 3.358,
    "args": {
      "External id": 173,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 92,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 92, "pid": 49625, "tid": 49625, "ts": 7253740876029.624,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876037.028, "dur": 3.220,
    "args": {
      "External id": 174,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 93,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 93, "pid": 49625, "tid": 49625, "ts": 7253740876037.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876043.629, "dur": 3.161,
    "args": {
      "External id": 175,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 94,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 94, "pid": 49625, "tid": 49625, "ts": 7253740876043.629,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876051.018, "dur": 2.935,
    "args": {
      "External id": 176,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 95,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 95, "pid": 49625, "tid": 49625, "ts": 7253740876051.018,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876058.867, "dur": 3.171,
    "args": {
      "External id": 177,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 96,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 96, "pid": 49625, "tid": 49625, "ts": 7253740876058.867,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876066.269, "dur": 3.866,
    "args": {
      "External id": 179,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 97,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 97, "pid": 49625, "tid": 49625, "ts": 7253740876066.269,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876073.457, "dur": 4.280,
    "args": {
      "External id": 180,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 98,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 98, "pid": 49625, "tid": 49625, "ts": 7253740876073.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876087.926, "dur": 6.633,
    "args": {
      "External id": 183,      "cid": 134, "correlation": 99, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 99, "pid": 49625, "tid": 49625, "ts": 7253740876087.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876118.962, "dur": 4.187,
    "args": {
      "External id": 190,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 100,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 100, "pid": 49625, "tid": 49625, "ts": 7253740876118.962,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876127.855, "dur": 3.151,
    "args": {
      "External id": 191,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 101,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 101, "pid": 49625, "tid": 49625, "ts": 7253740876127.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876135.178, "dur": 3.862,
    "args": {
      "External id": 192,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 102,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 102, "pid": 49625, "tid": 49625, "ts": 7253740876135.178,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876142.579, "dur": 3.931,
    "args": {
      "External id": 193,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 103,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 103, "pid": 49625, "tid": 49625, "ts": 7253740876142.579,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876150.698, "dur": 4.406,
    "args": {
      "External id": 194,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 104,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 104, "pid": 49625, "tid": 49625, "ts": 7253740876150.698,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876160.068, "dur": 3.955,
    "args": {
      "External id": 195,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 105,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 105, "pid": 49625, "tid": 49625, "ts": 7253740876160.068,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876168.121, "dur": 3.711,
    "args": {
      "External id": 197,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 106,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 106, "pid": 49625, "tid": 49625, "ts": 7253740876168.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876175.092, "dur": 3.657,
    "args": {
      "External id": 198,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 107,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 107, "pid": 49625, "tid": 49625, "ts": 7253740876175.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876188.806, "dur": 3.587,
    "args": {
      "External id": 201,      "cid": 134, "correlation": 108, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 108, "pid": 49625, "tid": 49625, "ts": 7253740876188.806,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876218.953, "dur": 3.407,
    "args": {
      "External id": 208,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 109,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 109, "pid": 49625, "tid": 49625, "ts": 7253740876218.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876227.038, "dur": 3.784,
    "args": {
      "External id": 209,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 110,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 110, "pid": 49625, "tid": 49625, "ts": 7253740876227.038,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876235.265, "dur": 5.009,
    "args": {
      "External id": 210,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 111,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 111, "pid": 49625, "tid": 49625, "ts": 7253740876235.265,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876243.831, "dur": 5.361,
    "args": {
      "External id": 211,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 112,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 112, "pid": 49625, "tid": 49625, "ts": 7253740876243.831,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876253.408, "dur": 4.497,
    "args": {
      "External id": 212,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 113,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 113, "pid": 49625, "tid": 49625, "ts": 7253740876253.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876263.049, "dur": 4.077,
    "args": {
      "External id": 213,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 114,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 114, "pid": 49625, "tid": 49625, "ts": 7253740876263.049,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876271.276, "dur": 3.726,
    "args": {
      "External id": 215,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 115,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 115, "pid": 49625, "tid": 49625, "ts": 7253740876271.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876278.372, "dur": 3.689,
    "args": {
      "External id": 216,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 116,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 116, "pid": 49625, "tid": 49625, "ts": 7253740876278.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876292.318, "dur": 4.920,
    "args": {
      "External id": 219,      "cid": 134, "correlation": 117, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 117, "pid": 49625, "tid": 49625, "ts": 7253740876292.318,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876322.369, "dur": 4.274,
    "args": {
      "External id": 226,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 118,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 118, "pid": 49625, "tid": 49625, "ts": 7253740876322.369,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876331.065, "dur": 3.933,
    "args": {
      "External id": 227,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 119,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 119, "pid": 49625, "tid": 49625, "ts": 7253740876331.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876339.610, "dur": 3.745,
    "args": {
      "External id": 228,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 120,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 120, "pid": 49625, "tid": 49625, "ts": 7253740876339.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876346.843, "dur": 3.618,
    "args": {
      "External id": 229,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 121,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 121, "pid": 49625, "tid": 49625, "ts": 7253740876346.843,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876354.588, "dur": 5.061,
    "args": {
      "External id": 230,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 122,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 122, "pid": 49625, "tid": 49625, "ts": 7253740876354.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876364.722, "dur": 3.944,
    "args": {
      "External id": 231,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 123,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 123, "pid": 49625, "tid": 49625, "ts": 7253740876364.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876372.781, "dur": 3.654,
    "args": {
      "External id": 233,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 124,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 124, "pid": 49625, "tid": 49625, "ts": 7253740876372.781,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876380.719, "dur": 3.548,
    "args": {
      "External id": 234,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 125,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 125, "pid": 49625, "tid": 49625, "ts": 7253740876380.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876394.483, "dur": 4.091,
    "args": {
      "External id": 237,      "cid": 134, "correlation": 126, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 126, "pid": 49625, "tid": 49625, "ts": 7253740876394.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876422.982, "dur": 3.769,
    "args": {
      "External id": 244,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 127,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 127, "pid": 49625, "tid": 49625, "ts": 7253740876422.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876431.120, "dur": 4.038,
    "args": {
      "External id": 245,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 128,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 128, "pid": 49625, "tid": 49625, "ts": 7253740876431.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876439.314, "dur": 4.088,
    "args": {
      "External id": 246,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 129,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 129, "pid": 49625, "tid": 49625, "ts": 7253740876439.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876448.254, "dur": 3.665,
    "args": {
      "External id": 247,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 130,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 130, "pid": 49625, "tid": 49625, "ts": 7253740876448.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876455.977, "dur": 3.874,
    "args": {
      "External id": 248,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 131,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 131, "pid": 49625, "tid": 49625, "ts": 7253740876455.977,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876464.819, "dur": 3.743,
    "args": {
      "External id": 249,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 132,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 132, "pid": 49625, "tid": 49625, "ts": 7253740876464.819,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876472.772, "dur": 3.669,
    "args": {
      "External id": 251,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 133,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 133, "pid": 49625, "tid": 49625, "ts": 7253740876472.772,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876479.821, "dur": 3.730,
    "args": {
      "External id": 252,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 134,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 134, "pid": 49625, "tid": 49625, "ts": 7253740876479.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876493.426, "dur": 3.375,
    "args": {
      "External id": 255,      "cid": 134, "correlation": 135, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 135, "pid": 49625, "tid": 49625, "ts": 7253740876493.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876521.445, "dur": 4.295,
    "args": {
      "External id": 262,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 136,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 136, "pid": 49625, "tid": 49625, "ts": 7253740876521.445,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876530.232, "dur": 3.701,
    "args": {
      "External id": 263,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 137,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 137, "pid": 49625, "tid": 49625, "ts": 7253740876530.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876538.136, "dur": 3.404,
    "args": {
      "External id": 264,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 138,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 138, "pid": 49625, "tid": 49625, "ts": 7253740876538.136,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876545.054, "dur": 3.867,
    "args": {
      "External id": 265,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 139,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 139, "pid": 49625, "tid": 49625, "ts": 7253740876545.054,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876552.988, "dur": 3.744,
    "args": {
      "External id": 266,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 140,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 140, "pid": 49625, "tid": 49625, "ts": 7253740876552.988,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876561.590, "dur": 3.115,
    "args": {
      "External id": 267,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 141,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 141, "pid": 49625, "tid": 49625, "ts": 7253740876561.590,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876568.824, "dur": 3.473,
    "args": {
      "External id": 269,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 142,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 142, "pid": 49625, "tid": 49625, "ts": 7253740876568.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876575.690, "dur": 3.579,
    "args": {
      "External id": 270,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 143,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 143, "pid": 49625, "tid": 49625, "ts": 7253740876575.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876589.119, "dur": 4.619,
    "args": {
      "External id": 273,      "cid": 134, "correlation": 144, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 144, "pid": 49625, "tid": 49625, "ts": 7253740876589.119,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876617.938, "dur": 4.193,
    "args": {
      "External id": 280,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 145,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 145, "pid": 49625, "tid": 49625, "ts": 7253740876617.938,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876626.846, "dur": 3.781,
    "args": {
      "External id": 281,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 146,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 146, "pid": 49625, "tid": 49625, "ts": 7253740876626.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876634.738, "dur": 3.928,
    "args": {
      "External id": 282,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 147,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 147, "pid": 49625, "tid": 49625, "ts": 7253740876634.738,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876642.104, "dur": 3.720,
    "args": {
      "External id": 283,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 148,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 148, "pid": 49625, "tid": 49625, "ts": 7253740876642.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876649.753, "dur": 4.778,
    "args": {
      "External id": 284,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 149,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 149, "pid": 49625, "tid": 49625, "ts": 7253740876649.753,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876659.301, "dur": 4.400,
    "args": {
      "External id": 285,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 150,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 150, "pid": 49625, "tid": 49625, "ts": 7253740876659.301,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876667.787, "dur": 3.850,
    "args": {
      "External id": 287,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 151,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 151, "pid": 49625, "tid": 49625, "ts": 7253740876667.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876674.881, "dur": 3.792,
    "args": {
      "External id": 288,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 152,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 152, "pid": 49625, "tid": 49625, "ts": 7253740876674.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876688.509, "dur": 4.522,
    "args": {
      "External id": 291,      "cid": 134, "correlation": 153, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 153, "pid": 49625, "tid": 49625, "ts": 7253740876688.509,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876717.076, "dur": 3.295,
    "args": {
      "External id": 298,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 154,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 154, "pid": 49625, "tid": 49625, "ts": 7253740876717.076,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876724.791, "dur": 3.117,
    "args": {
      "External id": 299,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 155,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 155, "pid": 49625, "tid": 49625, "ts": 7253740876724.791,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876731.904, "dur": 3.056,
    "args": {
      "External id": 300,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 156,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 156, "pid": 49625, "tid": 49625, "ts": 7253740876731.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876738.613, "dur": 4.658,
    "args": {
      "External id": 301,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 157,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 157, "pid": 49625, "tid": 49625, "ts": 7253740876738.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876747.270, "dur": 3.907,
    "args": {
      "External id": 302,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 158,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 158, "pid": 49625, "tid": 49625, "ts": 7253740876747.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876756.173, "dur": 3.724,
    "args": {
      "External id": 303,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 159,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 159, "pid": 49625, "tid": 49625, "ts": 7253740876756.173,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876763.950, "dur": 4.339,
    "args": {
      "External id": 305,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 160,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 160, "pid": 49625, "tid": 49625, "ts": 7253740876763.950,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876771.496, "dur": 3.808,
    "args": {
      "External id": 306,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 161,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 161, "pid": 49625, "tid": 49625, "ts": 7253740876771.496,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876785.129, "dur": 3.649,
    "args": {
      "External id": 309,      "cid": 134, "correlation": 162, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 162, "pid": 49625, "tid": 49625, "ts": 7253740876785.129,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876813.894, "dur": 4.659,
    "args": {
      "External id": 316,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 163,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 163, "pid": 49625, "tid": 49625, "ts": 7253740876813.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876823.064, "dur": 3.824,
    "args": {
      "External id": 317,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 164,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 164, "pid": 49625, "tid": 49625, "ts": 7253740876823.064,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876831.145, "dur": 3.754,
    "args": {
      "External id": 318,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 165,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 165, "pid": 49625, "tid": 49625, "ts": 7253740876831.145,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876838.263, "dur": 4.261,
    "args": {
      "External id": 319,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 166,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 166, "pid": 49625, "tid": 49625, "ts": 7253740876838.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876846.453, "dur": 4.076,
    "args": {
      "External id": 320,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 167,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 167, "pid": 49625, "tid": 49625, "ts": 7253740876846.453,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876855.397, "dur": 3.470,
    "args": {
      "External id": 321,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 168,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 168, "pid": 49625, "tid": 49625, "ts": 7253740876855.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876862.869, "dur": 4.006,
    "args": {
      "External id": 323,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 169,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 169, "pid": 49625, "tid": 49625, "ts": 7253740876862.869,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876870.156, "dur": 4.672,
    "args": {
      "External id": 324,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 170,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 170, "pid": 49625, "tid": 49625, "ts": 7253740876870.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740876884.573, "dur": 4.306,
    "args": {
      "External id": 327,      "cid": 134, "correlation": 171, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 171, "pid": 49625, "tid": 49625, "ts": 7253740876884.573,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876912.492, "dur": 31.355,
    "args": {
      "External id": 334,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 172,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 172, "pid": 49625, "tid": 49625, "ts": 7253740876912.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876948.627, "dur": 4.235,
    "args": {
      "External id": 335,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 173,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 173, "pid": 49625, "tid": 49625, "ts": 7253740876948.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876957.400, "dur": 4.466,
    "args": {
      "External id": 336,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 174,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 174, "pid": 49625, "tid": 49625, "ts": 7253740876957.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876965.471, "dur": 4.794,
    "args": {
      "External id": 337,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 175,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 175, "pid": 49625, "tid": 49625, "ts": 7253740876965.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876974.343, "dur": 4.663,
    "args": {
      "External id": 338,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 176,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 176, "pid": 49625, "tid": 49625, "ts": 7253740876974.343,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876983.935, "dur": 4.371,
    "args": {
      "External id": 339,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 177,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 177, "pid": 49625, "tid": 49625, "ts": 7253740876983.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876992.696, "dur": 3.550,
    "args": {
      "External id": 341,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 178,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 178, "pid": 49625, "tid": 49625, "ts": 7253740876992.696,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740876999.521, "dur": 4.217,
    "args": {
      "External id": 342,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 179,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 179, "pid": 49625, "tid": 49625, "ts": 7253740876999.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877014.415, "dur": 5.119,
    "args": {
      "External id": 345,      "cid": 134, "correlation": 180, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 180, "pid": 49625, "tid": 49625, "ts": 7253740877014.415,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877044.554, "dur": 4.689,
    "args": {
      "External id": 352,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 181,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 181, "pid": 49625, "tid": 49625, "ts": 7253740877044.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877054.303, "dur": 5.821,
    "args": {
      "External id": 353,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 182,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 182, "pid": 49625, "tid": 49625, "ts": 7253740877054.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877064.317, "dur": 3.470,
    "args": {
      "External id": 354,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 183,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 183, "pid": 49625, "tid": 49625, "ts": 7253740877064.317,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877071.235, "dur": 3.775,
    "args": {
      "External id": 355,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 184,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 184, "pid": 49625, "tid": 49625, "ts": 7253740877071.235,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877078.928, "dur": 4.199,
    "args": {
      "External id": 356,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 185,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 185, "pid": 49625, "tid": 49625, "ts": 7253740877078.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877088.467, "dur": 4.787,
    "args": {
      "External id": 357,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 186,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 186, "pid": 49625, "tid": 49625, "ts": 7253740877088.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877097.223, "dur": 4.049,
    "args": {
      "External id": 359,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 187,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 187, "pid": 49625, "tid": 49625, "ts": 7253740877097.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877104.580, "dur": 4.521,
    "args": {
      "External id": 360,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 188,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 188, "pid": 49625, "tid": 49625, "ts": 7253740877104.580,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877119.147, "dur": 4.904,
    "args": {
      "External id": 363,      "cid": 134, "correlation": 189, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 189, "pid": 49625, "tid": 49625, "ts": 7253740877119.147,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877148.324, "dur": 3.289,
    "args": {
      "External id": 370,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 190,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 190, "pid": 49625, "tid": 49625, "ts": 7253740877148.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877155.903, "dur": 4.037,
    "args": {
      "External id": 371,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 191,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 191, "pid": 49625, "tid": 49625, "ts": 7253740877155.903,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877163.999, "dur": 5.025,
    "args": {
      "External id": 372,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 192,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 192, "pid": 49625, "tid": 49625, "ts": 7253740877163.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877173.108, "dur": 4.459,
    "args": {
      "External id": 373,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 193,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 193, "pid": 49625, "tid": 49625, "ts": 7253740877173.108,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877181.603, "dur": 4.916,
    "args": {
      "External id": 374,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 194,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 194, "pid": 49625, "tid": 49625, "ts": 7253740877181.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877191.426, "dur": 3.903,
    "args": {
      "External id": 375,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 195,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 195, "pid": 49625, "tid": 49625, "ts": 7253740877191.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877199.309, "dur": 4.215,
    "args": {
      "External id": 377,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 196,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 196, "pid": 49625, "tid": 49625, "ts": 7253740877199.309,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877207.116, "dur": 3.584,
    "args": {
      "External id": 378,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 197,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 197, "pid": 49625, "tid": 49625, "ts": 7253740877207.116,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877220.397, "dur": 3.490,
    "args": {
      "External id": 381,      "cid": 134, "correlation": 198, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 198, "pid": 49625, "tid": 49625, "ts": 7253740877220.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877247.561, "dur": 4.820,
    "args": {
      "External id": 388,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 199,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 199, "pid": 49625, "tid": 49625, "ts": 7253740877247.561,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877257.414, "dur": 4.879,
    "args": {
      "External id": 389,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 200,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 200, "pid": 49625, "tid": 49625, "ts": 7253740877257.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877266.446, "dur": 3.765,
    "args": {
      "External id": 390,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 201,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 201, "pid": 49625, "tid": 49625, "ts": 7253740877266.446,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877273.660, "dur": 3.681,
    "args": {
      "External id": 391,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 202,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 202, "pid": 49625, "tid": 49625, "ts": 7253740877273.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877281.307, "dur": 4.807,
    "args": {
      "External id": 392,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 203,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 203, "pid": 49625, "tid": 49625, "ts": 7253740877281.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877291.118, "dur": 3.257,
    "args": {
      "External id": 393,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 204,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 204, "pid": 49625, "tid": 49625, "ts": 7253740877291.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877298.327, "dur": 3.942,
    "args": {
      "External id": 395,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 205,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 205, "pid": 49625, "tid": 49625, "ts": 7253740877298.327,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877305.543, "dur": 5.152,
    "args": {
      "External id": 396,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 206,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 206, "pid": 49625, "tid": 49625, "ts": 7253740877305.543,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877320.541, "dur": 4.476,
    "args": {
      "External id": 399,      "cid": 134, "correlation": 207, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 207, "pid": 49625, "tid": 49625, "ts": 7253740877320.541,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877348.533, "dur": 4.206,
    "args": {
      "External id": 406,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 208,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 208, "pid": 49625, "tid": 49625, "ts": 7253740877348.533,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877356.980, "dur": 3.551,
    "args": {
      "External id": 407,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 209,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 209, "pid": 49625, "tid": 49625, "ts": 7253740877356.980,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877364.578, "dur": 5.068,
    "args": {
      "External id": 408,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 210,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 210, "pid": 49625, "tid": 49625, "ts": 7253740877364.578,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877373.029, "dur": 4.646,
    "args": {
      "External id": 409,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 211,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 211, "pid": 49625, "tid": 49625, "ts": 7253740877373.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877381.705, "dur": 4.954,
    "args": {
      "External id": 410,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 212,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 212, "pid": 49625, "tid": 49625, "ts": 7253740877381.705,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877391.494, "dur": 4.653,
    "args": {
      "External id": 411,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 213,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 213, "pid": 49625, "tid": 49625, "ts": 7253740877391.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877400.323, "dur": 4.537,
    "args": {
      "External id": 413,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 214,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 214, "pid": 49625, "tid": 49625, "ts": 7253740877400.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877408.083, "dur": 3.826,
    "args": {
      "External id": 414,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 215,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 215, "pid": 49625, "tid": 49625, "ts": 7253740877408.083,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877421.748, "dur": 4.055,
    "args": {
      "External id": 417,      "cid": 134, "correlation": 216, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 216, "pid": 49625, "tid": 49625, "ts": 7253740877421.748,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877449.814, "dur": 4.934,
    "args": {
      "External id": 424,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 217,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 217, "pid": 49625, "tid": 49625, "ts": 7253740877449.814,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877459.254, "dur": 4.211,
    "args": {
      "External id": 425,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 218,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 218, "pid": 49625, "tid": 49625, "ts": 7253740877459.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877467.576, "dur": 3.739,
    "args": {
      "External id": 426,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 219,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 219, "pid": 49625, "tid": 49625, "ts": 7253740877467.576,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877474.790, "dur": 3.791,
    "args": {
      "External id": 427,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 220,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 220, "pid": 49625, "tid": 49625, "ts": 7253740877474.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877482.555, "dur": 3.649,
    "args": {
      "External id": 428,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 221,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 221, "pid": 49625, "tid": 49625, "ts": 7253740877482.555,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877491.136, "dur": 3.929,
    "args": {
      "External id": 429,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 222,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 222, "pid": 49625, "tid": 49625, "ts": 7253740877491.136,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877498.954, "dur": 4.416,
    "args": {
      "External id": 431,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 223,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 223, "pid": 49625, "tid": 49625, "ts": 7253740877498.954,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877506.645, "dur": 4.008,
    "args": {
      "External id": 432,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 224,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 224, "pid": 49625, "tid": 49625, "ts": 7253740877506.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877520.466, "dur": 3.837,
    "args": {
      "External id": 435,      "cid": 134, "correlation": 225, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 225, "pid": 49625, "tid": 49625, "ts": 7253740877520.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877548.274, "dur": 5.076,
    "args": {
      "External id": 442,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 226,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 226, "pid": 49625, "tid": 49625, "ts": 7253740877548.274,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877557.925, "dur": 3.408,
    "args": {
      "External id": 443,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 227,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 227, "pid": 49625, "tid": 49625, "ts": 7253740877557.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877565.463, "dur": 3.614,
    "args": {
      "External id": 444,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 228,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 228, "pid": 49625, "tid": 49625, "ts": 7253740877565.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877572.513, "dur": 4.505,
    "args": {
      "External id": 445,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 229,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 229, "pid": 49625, "tid": 49625, "ts": 7253740877572.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877581.056, "dur": 4.021,
    "args": {
      "External id": 446,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 230,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 230, "pid": 49625, "tid": 49625, "ts": 7253740877581.056,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877589.926, "dur": 3.665,
    "args": {
      "External id": 447,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 231,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 231, "pid": 49625, "tid": 49625, "ts": 7253740877589.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877597.549, "dur": 3.525,
    "args": {
      "External id": 449,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 232,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 232, "pid": 49625, "tid": 49625, "ts": 7253740877597.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877605.359, "dur": 3.844,
    "args": {
      "External id": 450,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 233,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 233, "pid": 49625, "tid": 49625, "ts": 7253740877605.359,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877618.940, "dur": 3.930,
    "args": {
      "External id": 453,      "cid": 134, "correlation": 234, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 234, "pid": 49625, "tid": 49625, "ts": 7253740877618.940,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877646.859, "dur": 3.922,
    "args": {
      "External id": 460,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 235,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 235, "pid": 49625, "tid": 49625, "ts": 7253740877646.859,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877655.212, "dur": 4.379,
    "args": {
      "External id": 461,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 236,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 236, "pid": 49625, "tid": 49625, "ts": 7253740877655.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877663.688, "dur": 4.087,
    "args": {
      "External id": 462,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 237,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 237, "pid": 49625, "tid": 49625, "ts": 7253740877663.688,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877671.229, "dur": 3.899,
    "args": {
      "External id": 463,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 238,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 238, "pid": 49625, "tid": 49625, "ts": 7253740877671.229,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877679.166, "dur": 3.678,
    "args": {
      "External id": 464,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 239,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 239, "pid": 49625, "tid": 49625, "ts": 7253740877679.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877687.875, "dur": 3.935,
    "args": {
      "External id": 465,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 240,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 240, "pid": 49625, "tid": 49625, "ts": 7253740877687.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877696.410, "dur": 3.707,
    "args": {
      "External id": 467,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 241,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 241, "pid": 49625, "tid": 49625, "ts": 7253740877696.410,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877703.410, "dur": 3.040,
    "args": {
      "External id": 468,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 242,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 242, "pid": 49625, "tid": 49625, "ts": 7253740877703.410,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877716.299, "dur": 3.491,
    "args": {
      "External id": 471,      "cid": 134, "correlation": 243, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 243, "pid": 49625, "tid": 49625, "ts": 7253740877716.299,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877750.527, "dur": 3.330,
    "args": {
      "External id": 478,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 244,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 244, "pid": 49625, "tid": 49625, "ts": 7253740877750.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877758.346, "dur": 4.214,
    "args": {
      "External id": 479,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 245,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 245, "pid": 49625, "tid": 49625, "ts": 7253740877758.346,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877766.851, "dur": 4.718,
    "args": {
      "External id": 480,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 246,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 246, "pid": 49625, "tid": 49625, "ts": 7253740877766.851,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877775.049, "dur": 5.930,
    "args": {
      "External id": 481,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 247,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 247, "pid": 49625, "tid": 49625, "ts": 7253740877775.049,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877785.264, "dur": 5.567,
    "args": {
      "External id": 482,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 248,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 248, "pid": 49625, "tid": 49625, "ts": 7253740877785.264,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877795.886, "dur": 3.442,
    "args": {
      "External id": 483,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 249,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 249, "pid": 49625, "tid": 49625, "ts": 7253740877795.886,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877803.389, "dur": 4.336,
    "args": {
      "External id": 485,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 250,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 250, "pid": 49625, "tid": 49625, "ts": 7253740877803.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877811.329, "dur": 4.539,
    "args": {
      "External id": 486,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 251,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 251, "pid": 49625, "tid": 49625, "ts": 7253740877811.329,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877826.105, "dur": 4.333,
    "args": {
      "External id": 489,      "cid": 134, "correlation": 252, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 252, "pid": 49625, "tid": 49625, "ts": 7253740877826.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877855.317, "dur": 3.380,
    "args": {
      "External id": 496,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 253,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 253, "pid": 49625, "tid": 49625, "ts": 7253740877855.317,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877863.438, "dur": 5.547,
    "args": {
      "External id": 497,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 254,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 254, "pid": 49625, "tid": 49625, "ts": 7253740877863.438,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877873.270, "dur": 4.233,
    "args": {
      "External id": 498,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 255,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 255, "pid": 49625, "tid": 49625, "ts": 7253740877873.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877880.987, "dur": 3.662,
    "args": {
      "External id": 499,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 256,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 256, "pid": 49625, "tid": 49625, "ts": 7253740877880.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877888.653, "dur": 4.901,
    "args": {
      "External id": 500,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 257,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 257, "pid": 49625, "tid": 49625, "ts": 7253740877888.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877900.552, "dur": 3.830,
    "args": {
      "External id": 501,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 258,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 258, "pid": 49625, "tid": 49625, "ts": 7253740877900.552,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877908.287, "dur": 4.166,
    "args": {
      "External id": 503,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 259,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 259, "pid": 49625, "tid": 49625, "ts": 7253740877908.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877915.684, "dur": 5.361,
    "args": {
      "External id": 504,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 260,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 260, "pid": 49625, "tid": 49625, "ts": 7253740877915.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740877931.307, "dur": 3.829,
    "args": {
      "External id": 507,      "cid": 134, "correlation": 261, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 261, "pid": 49625, "tid": 49625, "ts": 7253740877931.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877986.858, "dur": 4.833,
    "args": {
      "External id": 514,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 262,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 262, "pid": 49625, "tid": 49625, "ts": 7253740877986.858,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740877996.164, "dur": 3.572,
    "args": {
      "External id": 515,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 263,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 263, "pid": 49625, "tid": 49625, "ts": 7253740877996.164,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878003.919, "dur": 3.844,
    "args": {
      "External id": 516,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 264,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 264, "pid": 49625, "tid": 49625, "ts": 7253740878003.919,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878011.350, "dur": 4.173,
    "args": {
      "External id": 517,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 265,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 265, "pid": 49625, "tid": 49625, "ts": 7253740878011.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878019.893, "dur": 4.158,
    "args": {
      "External id": 518,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 266,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 266, "pid": 49625, "tid": 49625, "ts": 7253740878019.893,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878029.155, "dur": 4.332,
    "args": {
      "External id": 519,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 267,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 267, "pid": 49625, "tid": 49625, "ts": 7253740878029.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878037.457, "dur": 3.944,
    "args": {
      "External id": 521,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 268,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 268, "pid": 49625, "tid": 49625, "ts": 7253740878037.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878044.669, "dur": 4.116,
    "args": {
      "External id": 522,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 269,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 269, "pid": 49625, "tid": 49625, "ts": 7253740878044.669,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878058.695, "dur": 4.498,
    "args": {
      "External id": 525,      "cid": 134, "correlation": 270, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 270, "pid": 49625, "tid": 49625, "ts": 7253740878058.695,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878132.046, "dur": 4.428,
    "args": {
      "External id": 532,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 271,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 271, "pid": 49625, "tid": 49625, "ts": 7253740878132.046,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878143.085, "dur": 7.042,
    "args": {
      "External id": 533,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 272,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 272, "pid": 49625, "tid": 49625, "ts": 7253740878143.085,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878155.631, "dur": 4.702,
    "args": {
      "External id": 534,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 273,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 273, "pid": 49625, "tid": 49625, "ts": 7253740878155.631,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878164.153, "dur": 4.244,
    "args": {
      "External id": 535,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 274,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 274, "pid": 49625, "tid": 49625, "ts": 7253740878164.153,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878173.137, "dur": 4.597,
    "args": {
      "External id": 536,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 275,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 275, "pid": 49625, "tid": 49625, "ts": 7253740878173.137,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878183.259, "dur": 5.187,
    "args": {
      "External id": 537,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 276,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 276, "pid": 49625, "tid": 49625, "ts": 7253740878183.259,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878192.946, "dur": 4.110,
    "args": {
      "External id": 539,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 277,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 277, "pid": 49625, "tid": 49625, "ts": 7253740878192.946,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878200.460, "dur": 4.987,
    "args": {
      "External id": 540,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 278,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 278, "pid": 49625, "tid": 49625, "ts": 7253740878200.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878218.711, "dur": 4.643,
    "args": {
      "External id": 543,      "cid": 134, "correlation": 279, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 279, "pid": 49625, "tid": 49625, "ts": 7253740878218.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878251.030, "dur": 3.558,
    "args": {
      "External id": 550,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 280,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 280, "pid": 49625, "tid": 49625, "ts": 7253740878251.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878259.498, "dur": 3.291,
    "args": {
      "External id": 551,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 281,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 281, "pid": 49625, "tid": 49625, "ts": 7253740878259.498,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878267.027, "dur": 4.395,
    "args": {
      "External id": 552,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 282,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 282, "pid": 49625, "tid": 49625, "ts": 7253740878267.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878275.128, "dur": 3.894,
    "args": {
      "External id": 553,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 283,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 283, "pid": 49625, "tid": 49625, "ts": 7253740878275.128,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878283.151, "dur": 4.057,
    "args": {
      "External id": 554,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 284,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 284, "pid": 49625, "tid": 49625, "ts": 7253740878283.151,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878292.409, "dur": 3.742,
    "args": {
      "External id": 555,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 285,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 285, "pid": 49625, "tid": 49625, "ts": 7253740878292.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878300.290, "dur": 4.021,
    "args": {
      "External id": 557,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 286,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 286, "pid": 49625, "tid": 49625, "ts": 7253740878300.290,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878307.692, "dur": 5.039,
    "args": {
      "External id": 558,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 287,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 287, "pid": 49625, "tid": 49625, "ts": 7253740878307.692,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878323.243, "dur": 3.521,
    "args": {
      "External id": 561,      "cid": 134, "correlation": 288, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 288, "pid": 49625, "tid": 49625, "ts": 7253740878323.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878352.720, "dur": 4.192,
    "args": {
      "External id": 568,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 289,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 289, "pid": 49625, "tid": 49625, "ts": 7253740878352.720,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878361.610, "dur": 3.815,
    "args": {
      "External id": 569,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 290,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 290, "pid": 49625, "tid": 49625, "ts": 7253740878361.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878369.681, "dur": 3.607,
    "args": {
      "External id": 570,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 291,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 291, "pid": 49625, "tid": 49625, "ts": 7253740878369.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878376.709, "dur": 3.114,
    "args": {
      "External id": 571,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 292,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 292, "pid": 49625, "tid": 49625, "ts": 7253740878376.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878383.823, "dur": 4.133,
    "args": {
      "External id": 572,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 293,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 293, "pid": 49625, "tid": 49625, "ts": 7253740878383.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878392.868, "dur": 4.004,
    "args": {
      "External id": 573,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 294,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 294, "pid": 49625, "tid": 49625, "ts": 7253740878392.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878400.855, "dur": 4.144,
    "args": {
      "External id": 575,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 295,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 295, "pid": 49625, "tid": 49625, "ts": 7253740878400.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878408.332, "dur": 4.460,
    "args": {
      "External id": 576,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 296,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 296, "pid": 49625, "tid": 49625, "ts": 7253740878408.332,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878422.948, "dur": 4.416,
    "args": {
      "External id": 579,      "cid": 134, "correlation": 297, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 297, "pid": 49625, "tid": 49625, "ts": 7253740878422.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878452.850, "dur": 3.628,
    "args": {
      "External id": 586,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 298,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 298, "pid": 49625, "tid": 49625, "ts": 7253740878452.850,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878460.757, "dur": 3.993,
    "args": {
      "External id": 587,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 299,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 299, "pid": 49625, "tid": 49625, "ts": 7253740878460.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878469.003, "dur": 4.451,
    "args": {
      "External id": 588,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 300,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 300, "pid": 49625, "tid": 49625, "ts": 7253740878469.003,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878476.870, "dur": 3.803,
    "args": {
      "External id": 589,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 301,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 301, "pid": 49625, "tid": 49625, "ts": 7253740878476.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878484.803, "dur": 3.893,
    "args": {
      "External id": 590,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 302,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 302, "pid": 49625, "tid": 49625, "ts": 7253740878484.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878493.572, "dur": 3.724,
    "args": {
      "External id": 591,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 303,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 303, "pid": 49625, "tid": 49625, "ts": 7253740878493.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878501.256, "dur": 3.773,
    "args": {
      "External id": 593,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 304,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 304, "pid": 49625, "tid": 49625, "ts": 7253740878501.256,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878508.422, "dur": 3.535,
    "args": {
      "External id": 594,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 305,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 305, "pid": 49625, "tid": 49625, "ts": 7253740878508.422,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878522.230, "dur": 3.513,
    "args": {
      "External id": 597,      "cid": 134, "correlation": 306, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 306, "pid": 49625, "tid": 49625, "ts": 7253740878522.230,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878550.582, "dur": 4.307,
    "args": {
      "External id": 604,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 307,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 307, "pid": 49625, "tid": 49625, "ts": 7253740878550.582,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878559.402, "dur": 3.812,
    "args": {
      "External id": 605,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 308,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 308, "pid": 49625, "tid": 49625, "ts": 7253740878559.402,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878567.399, "dur": 3.558,
    "args": {
      "External id": 606,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 309,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 309, "pid": 49625, "tid": 49625, "ts": 7253740878567.399,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878574.743, "dur": 3.605,
    "args": {
      "External id": 607,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 310,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 310, "pid": 49625, "tid": 49625, "ts": 7253740878574.743,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878582.339, "dur": 3.961,
    "args": {
      "External id": 608,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 311,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 311, "pid": 49625, "tid": 49625, "ts": 7253740878582.339,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878591.211, "dur": 3.659,
    "args": {
      "External id": 609,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 312,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 312, "pid": 49625, "tid": 49625, "ts": 7253740878591.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878598.862, "dur": 4.242,
    "args": {
      "External id": 611,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 313,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 313, "pid": 49625, "tid": 49625, "ts": 7253740878598.862,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878606.380, "dur": 5.232,
    "args": {
      "External id": 612,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 314,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 314, "pid": 49625, "tid": 49625, "ts": 7253740878606.380,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878621.840, "dur": 4.527,
    "args": {
      "External id": 615,      "cid": 134, "correlation": 315, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 315, "pid": 49625, "tid": 49625, "ts": 7253740878621.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878651.038, "dur": 4.156,
    "args": {
      "External id": 622,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 316,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 316, "pid": 49625, "tid": 49625, "ts": 7253740878651.038,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878659.644, "dur": 3.842,
    "args": {
      "External id": 623,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 317,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 317, "pid": 49625, "tid": 49625, "ts": 7253740878659.644,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878667.581, "dur": 3.684,
    "args": {
      "External id": 624,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 318,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 318, "pid": 49625, "tid": 49625, "ts": 7253740878667.581,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878674.706, "dur": 3.938,
    "args": {
      "External id": 625,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 319,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 319, "pid": 49625, "tid": 49625, "ts": 7253740878674.706,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878682.845, "dur": 4.067,
    "args": {
      "External id": 626,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 320,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 320, "pid": 49625, "tid": 49625, "ts": 7253740878682.845,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878691.868, "dur": 4.174,
    "args": {
      "External id": 627,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 321,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 321, "pid": 49625, "tid": 49625, "ts": 7253740878691.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878700.189, "dur": 3.752,
    "args": {
      "External id": 629,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 322,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 322, "pid": 49625, "tid": 49625, "ts": 7253740878700.189,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878707.308, "dur": 4.267,
    "args": {
      "External id": 630,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 323,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 323, "pid": 49625, "tid": 49625, "ts": 7253740878707.308,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878721.520, "dur": 3.976,
    "args": {
      "External id": 633,      "cid": 134, "correlation": 324, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 324, "pid": 49625, "tid": 49625, "ts": 7253740878721.520,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878750.406, "dur": 3.275,
    "args": {
      "External id": 640,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 325,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 325, "pid": 49625, "tid": 49625, "ts": 7253740878750.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878758.114, "dur": 3.331,
    "args": {
      "External id": 641,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 326,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 326, "pid": 49625, "tid": 49625, "ts": 7253740878758.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878765.558, "dur": 3.233,
    "args": {
      "External id": 642,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 327,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 327, "pid": 49625, "tid": 49625, "ts": 7253740878765.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878772.150, "dur": 4.195,
    "args": {
      "External id": 643,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 328,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 328, "pid": 49625, "tid": 49625, "ts": 7253740878772.150,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878780.472, "dur": 4.032,
    "args": {
      "External id": 644,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 329,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 329, "pid": 49625, "tid": 49625, "ts": 7253740878780.472,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878789.428, "dur": 3.611,
    "args": {
      "External id": 645,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 330,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 330, "pid": 49625, "tid": 49625, "ts": 7253740878789.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878797.146, "dur": 3.538,
    "args": {
      "External id": 647,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 331,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 331, "pid": 49625, "tid": 49625, "ts": 7253740878797.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878803.918, "dur": 3.640,
    "args": {
      "External id": 648,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 332,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 332, "pid": 49625, "tid": 49625, "ts": 7253740878803.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878817.299, "dur": 4.107,
    "args": {
      "External id": 651,      "cid": 134, "correlation": 333, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 333, "pid": 49625, "tid": 49625, "ts": 7253740878817.299,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878845.378, "dur": 3.079,
    "args": {
      "External id": 658,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 334,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 334, "pid": 49625, "tid": 49625, "ts": 7253740878845.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878852.673, "dur": 4.263,
    "args": {
      "External id": 659,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 335,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 335, "pid": 49625, "tid": 49625, "ts": 7253740878852.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878860.985, "dur": 3.988,
    "args": {
      "External id": 660,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 336,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 336, "pid": 49625, "tid": 49625, "ts": 7253740878860.985,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878868.442, "dur": 3.564,
    "args": {
      "External id": 661,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 337,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 337, "pid": 49625, "tid": 49625, "ts": 7253740878868.442,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878876.114, "dur": 3.864,
    "args": {
      "External id": 662,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 338,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 338, "pid": 49625, "tid": 49625, "ts": 7253740878876.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878884.937, "dur": 3.466,
    "args": {
      "External id": 663,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 339,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 339, "pid": 49625, "tid": 49625, "ts": 7253740878884.937,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878892.385, "dur": 4.273,
    "args": {
      "External id": 665,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 340,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 340, "pid": 49625, "tid": 49625, "ts": 7253740878892.385,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878899.964, "dur": 3.360,
    "args": {
      "External id": 666,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 341,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 341, "pid": 49625, "tid": 49625, "ts": 7253740878899.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740878913.415, "dur": 3.949,
    "args": {
      "External id": 669,      "cid": 134, "correlation": 342, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 342, "pid": 49625, "tid": 49625, "ts": 7253740878913.415,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878941.619, "dur": 4.392,
    "args": {
      "External id": 676,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 343,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 343, "pid": 49625, "tid": 49625, "ts": 7253740878941.619,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878950.463, "dur": 3.688,
    "args": {
      "External id": 677,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 344,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 344, "pid": 49625, "tid": 49625, "ts": 7253740878950.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878959.354, "dur": 3.649,
    "args": {
      "External id": 678,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 345,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 345, "pid": 49625, "tid": 49625, "ts": 7253740878959.354,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878966.400, "dur": 3.416,
    "args": {
      "External id": 679,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 346,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 346, "pid": 49625, "tid": 49625, "ts": 7253740878966.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878973.793, "dur": 3.503,
    "args": {
      "External id": 680,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 347,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 347, "pid": 49625, "tid": 49625, "ts": 7253740878973.793,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878982.252, "dur": 3.782,
    "args": {
      "External id": 681,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 348,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 348, "pid": 49625, "tid": 49625, "ts": 7253740878982.252,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878990.032, "dur": 3.928,
    "args": {
      "External id": 683,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 349,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 349, "pid": 49625, "tid": 49625, "ts": 7253740878990.032,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740878997.147, "dur": 4.140,
    "args": {
      "External id": 684,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 350,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 350, "pid": 49625, "tid": 49625, "ts": 7253740878997.147,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879011.148, "dur": 4.288,
    "args": {
      "External id": 687,      "cid": 134, "correlation": 351, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 351, "pid": 49625, "tid": 49625, "ts": 7253740879011.148,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879039.685, "dur": 4.528,
    "args": {
      "External id": 694,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 352,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 352, "pid": 49625, "tid": 49625, "ts": 7253740879039.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879048.691, "dur": 2.892,
    "args": {
      "External id": 695,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 353,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 353, "pid": 49625, "tid": 49625, "ts": 7253740879048.691,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879055.734, "dur": 3.779,
    "args": {
      "External id": 696,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 354,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 354, "pid": 49625, "tid": 49625, "ts": 7253740879055.734,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879062.983, "dur": 3.986,
    "args": {
      "External id": 697,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 355,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 355, "pid": 49625, "tid": 49625, "ts": 7253740879062.983,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879071.026, "dur": 4.512,
    "args": {
      "External id": 698,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 356,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 356, "pid": 49625, "tid": 49625, "ts": 7253740879071.026,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879081.319, "dur": 3.735,
    "args": {
      "External id": 699,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 357,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 357, "pid": 49625, "tid": 49625, "ts": 7253740879081.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879089.015, "dur": 3.599,
    "args": {
      "External id": 701,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 358,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 358, "pid": 49625, "tid": 49625, "ts": 7253740879089.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879095.802, "dur": 4.036,
    "args": {
      "External id": 702,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 359,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 359, "pid": 49625, "tid": 49625, "ts": 7253740879095.802,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879109.808, "dur": 4.640,
    "args": {
      "External id": 705,      "cid": 134, "correlation": 360, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 360, "pid": 49625, "tid": 49625, "ts": 7253740879109.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879140.498, "dur": 3.599,
    "args": {
      "External id": 712,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 361,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 361, "pid": 49625, "tid": 49625, "ts": 7253740879140.498,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879148.258, "dur": 4.997,
    "args": {
      "External id": 713,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 362,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 362, "pid": 49625, "tid": 49625, "ts": 7253740879148.258,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879157.240, "dur": 3.786,
    "args": {
      "External id": 714,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 363,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 363, "pid": 49625, "tid": 49625, "ts": 7253740879157.240,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879164.449, "dur": 3.660,
    "args": {
      "External id": 715,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 364,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 364, "pid": 49625, "tid": 49625, "ts": 7253740879164.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879172.298, "dur": 3.744,
    "args": {
      "External id": 716,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 365,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 365, "pid": 49625, "tid": 49625, "ts": 7253740879172.298,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879181.166, "dur": 3.843,
    "args": {
      "External id": 717,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 366,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 366, "pid": 49625, "tid": 49625, "ts": 7253740879181.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879188.975, "dur": 3.713,
    "args": {
      "External id": 719,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 367,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 367, "pid": 49625, "tid": 49625, "ts": 7253740879188.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879195.846, "dur": 3.944,
    "args": {
      "External id": 720,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 368,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 368, "pid": 49625, "tid": 49625, "ts": 7253740879195.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879209.668, "dur": 4.233,
    "args": {
      "External id": 723,      "cid": 134, "correlation": 369, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 369, "pid": 49625, "tid": 49625, "ts": 7253740879209.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879237.290, "dur": 3.561,
    "args": {
      "External id": 730,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 370,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 370, "pid": 49625, "tid": 49625, "ts": 7253740879237.290,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879244.998, "dur": 4.405,
    "args": {
      "External id": 731,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 371,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 371, "pid": 49625, "tid": 49625, "ts": 7253740879244.998,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879253.447, "dur": 4.214,
    "args": {
      "External id": 732,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 372,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 372, "pid": 49625, "tid": 49625, "ts": 7253740879253.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879261.080, "dur": 16.070,
    "args": {
      "External id": 733,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 373,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 373, "pid": 49625, "tid": 49625, "ts": 7253740879261.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879281.535, "dur": 3.791,
    "args": {
      "External id": 734,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 374,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 374, "pid": 49625, "tid": 49625, "ts": 7253740879281.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879290.616, "dur": 4.344,
    "args": {
      "External id": 735,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 375,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 375, "pid": 49625, "tid": 49625, "ts": 7253740879290.616,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879298.956, "dur": 3.632,
    "args": {
      "External id": 737,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 376,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 376, "pid": 49625, "tid": 49625, "ts": 7253740879298.956,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879311.251, "dur": 4.296,
    "args": {
      "External id": 738,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 377,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 377, "pid": 49625, "tid": 49625, "ts": 7253740879311.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879325.684, "dur": 4.044,
    "args": {
      "External id": 741,      "cid": 134, "correlation": 378, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 378, "pid": 49625, "tid": 49625, "ts": 7253740879325.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879353.765, "dur": 4.923,
    "args": {
      "External id": 748,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 379,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 379, "pid": 49625, "tid": 49625, "ts": 7253740879353.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879363.679, "dur": 3.903,
    "args": {
      "External id": 749,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 380,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 380, "pid": 49625, "tid": 49625, "ts": 7253740879363.679,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879371.618, "dur": 4.309,
    "args": {
      "External id": 750,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 381,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 381, "pid": 49625, "tid": 49625, "ts": 7253740879371.618,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879379.407, "dur": 3.617,
    "args": {
      "External id": 751,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 382,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 382, "pid": 49625, "tid": 49625, "ts": 7253740879379.407,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879387.167, "dur": 5.096,
    "args": {
      "External id": 752,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 383,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 383, "pid": 49625, "tid": 49625, "ts": 7253740879387.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879397.270, "dur": 3.995,
    "args": {
      "External id": 753,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 384,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 384, "pid": 49625, "tid": 49625, "ts": 7253740879397.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879405.154, "dur": 4.432,
    "args": {
      "External id": 755,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 385,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 385, "pid": 49625, "tid": 49625, "ts": 7253740879405.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879413.012, "dur": 4.974,
    "args": {
      "External id": 756,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 386,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 386, "pid": 49625, "tid": 49625, "ts": 7253740879413.012,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879427.889, "dur": 5.485,
    "args": {
      "External id": 759,      "cid": 134, "correlation": 387, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 387, "pid": 49625, "tid": 49625, "ts": 7253740879427.889,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879456.727, "dur": 3.265,
    "args": {
      "External id": 766,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 388,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 388, "pid": 49625, "tid": 49625, "ts": 7253740879456.727,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879464.241, "dur": 5.100,
    "args": {
      "External id": 767,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 389,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 389, "pid": 49625, "tid": 49625, "ts": 7253740879464.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879473.646, "dur": 4.183,
    "args": {
      "External id": 768,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 390,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 390, "pid": 49625, "tid": 49625, "ts": 7253740879473.646,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879481.210, "dur": 3.759,
    "args": {
      "External id": 769,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 391,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 391, "pid": 49625, "tid": 49625, "ts": 7253740879481.210,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879489.043, "dur": 4.060,
    "args": {
      "External id": 770,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 392,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 392, "pid": 49625, "tid": 49625, "ts": 7253740879489.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879497.981, "dur": 4.006,
    "args": {
      "External id": 771,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 393,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 393, "pid": 49625, "tid": 49625, "ts": 7253740879497.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879505.815, "dur": 3.920,
    "args": {
      "External id": 773,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 394,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 394, "pid": 49625, "tid": 49625, "ts": 7253740879505.815,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879512.884, "dur": 3.367,
    "args": {
      "External id": 774,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 395,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 395, "pid": 49625, "tid": 49625, "ts": 7253740879512.884,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879526.022, "dur": 3.426,
    "args": {
      "External id": 777,      "cid": 134, "correlation": 396, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 396, "pid": 49625, "tid": 49625, "ts": 7253740879526.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879553.548, "dur": 4.311,
    "args": {
      "External id": 784,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 397,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 397, "pid": 49625, "tid": 49625, "ts": 7253740879553.548,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879562.251, "dur": 4.112,
    "args": {
      "External id": 785,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 398,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 398, "pid": 49625, "tid": 49625, "ts": 7253740879562.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879570.559, "dur": 3.627,
    "args": {
      "External id": 786,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 399,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 399, "pid": 49625, "tid": 49625, "ts": 7253740879570.559,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879577.626, "dur": 3.994,
    "args": {
      "External id": 787,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 400,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 400, "pid": 49625, "tid": 49625, "ts": 7253740879577.626,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879585.658, "dur": 3.877,
    "args": {
      "External id": 788,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 401,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 401, "pid": 49625, "tid": 49625, "ts": 7253740879585.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879594.497, "dur": 3.213,
    "args": {
      "External id": 789,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 402,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 402, "pid": 49625, "tid": 49625, "ts": 7253740879594.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879601.757, "dur": 3.779,
    "args": {
      "External id": 791,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 403,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 403, "pid": 49625, "tid": 49625, "ts": 7253740879601.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879608.840, "dur": 4.293,
    "args": {
      "External id": 792,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 404,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 404, "pid": 49625, "tid": 49625, "ts": 7253740879608.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879623.179, "dur": 4.083,
    "args": {
      "External id": 795,      "cid": 134, "correlation": 405, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 405, "pid": 49625, "tid": 49625, "ts": 7253740879623.179,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879650.669, "dur": 4.373,
    "args": {
      "External id": 802,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 406,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 406, "pid": 49625, "tid": 49625, "ts": 7253740879650.669,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879659.486, "dur": 3.530,
    "args": {
      "External id": 803,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 407,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 407, "pid": 49625, "tid": 49625, "ts": 7253740879659.486,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879667.028, "dur": 3.673,
    "args": {
      "External id": 804,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 408,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 408, "pid": 49625, "tid": 49625, "ts": 7253740879667.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879674.133, "dur": 3.260,
    "args": {
      "External id": 805,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 409,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 409, "pid": 49625, "tid": 49625, "ts": 7253740879674.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879681.437, "dur": 4.554,
    "args": {
      "External id": 806,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 410,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 410, "pid": 49625, "tid": 49625, "ts": 7253740879681.437,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879690.943, "dur": 4.433,
    "args": {
      "External id": 807,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 411,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 411, "pid": 49625, "tid": 49625, "ts": 7253740879690.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879699.310, "dur": 3.564,
    "args": {
      "External id": 809,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 412,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 412, "pid": 49625, "tid": 49625, "ts": 7253740879699.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879706.109, "dur": 4.089,
    "args": {
      "External id": 810,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 413,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 413, "pid": 49625, "tid": 49625, "ts": 7253740879706.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879720.060, "dur": 4.293,
    "args": {
      "External id": 813,      "cid": 134, "correlation": 414, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 414, "pid": 49625, "tid": 49625, "ts": 7253740879720.060,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879747.455, "dur": 4.135,
    "args": {
      "External id": 820,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 415,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 415, "pid": 49625, "tid": 49625, "ts": 7253740879747.455,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879756.201, "dur": 4.496,
    "args": {
      "External id": 821,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 416,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 416, "pid": 49625, "tid": 49625, "ts": 7253740879756.201,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879764.881, "dur": 3.341,
    "args": {
      "External id": 822,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 417,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 417, "pid": 49625, "tid": 49625, "ts": 7253740879764.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879771.568, "dur": 3.857,
    "args": {
      "External id": 823,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 418,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 418, "pid": 49625, "tid": 49625, "ts": 7253740879771.568,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879779.408, "dur": 4.346,
    "args": {
      "External id": 824,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 419,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 419, "pid": 49625, "tid": 49625, "ts": 7253740879779.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879788.814, "dur": 3.812,
    "args": {
      "External id": 825,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 420,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 420, "pid": 49625, "tid": 49625, "ts": 7253740879788.814,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879796.509, "dur": 3.708,
    "args": {
      "External id": 827,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 421,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 421, "pid": 49625, "tid": 49625, "ts": 7253740879796.509,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879803.435, "dur": 4.394,
    "args": {
      "External id": 828,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 422,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 422, "pid": 49625, "tid": 49625, "ts": 7253740879803.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879817.488, "dur": 4.360,
    "args": {
      "External id": 831,      "cid": 134, "correlation": 423, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 423, "pid": 49625, "tid": 49625, "ts": 7253740879817.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879844.925, "dur": 3.252,
    "args": {
      "External id": 838,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 424,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 424, "pid": 49625, "tid": 49625, "ts": 7253740879844.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879852.464, "dur": 3.225,
    "args": {
      "External id": 839,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 425,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 425, "pid": 49625, "tid": 49625, "ts": 7253740879852.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879859.584, "dur": 3.147,
    "args": {
      "External id": 840,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 426,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 426, "pid": 49625, "tid": 49625, "ts": 7253740879859.584,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879866.169, "dur": 4.086,
    "args": {
      "External id": 841,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 427,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 427, "pid": 49625, "tid": 49625, "ts": 7253740879866.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879874.386, "dur": 3.703,
    "args": {
      "External id": 842,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 428,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 428, "pid": 49625, "tid": 49625, "ts": 7253740879874.386,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879882.962, "dur": 3.637,
    "args": {
      "External id": 843,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 429,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 429, "pid": 49625, "tid": 49625, "ts": 7253740879882.962,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879890.542, "dur": 3.698,
    "args": {
      "External id": 845,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 430,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 430, "pid": 49625, "tid": 49625, "ts": 7253740879890.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879897.454, "dur": 3.429,
    "args": {
      "External id": 846,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 431,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 431, "pid": 49625, "tid": 49625, "ts": 7253740879897.454,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740879910.507, "dur": 3.950,
    "args": {
      "External id": 849,      "cid": 134, "correlation": 432, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 432, "pid": 49625, "tid": 49625, "ts": 7253740879910.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879937.553, "dur": 3.666,
    "args": {
      "External id": 856,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 433,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 433, "pid": 49625, "tid": 49625, "ts": 7253740879937.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879945.459, "dur": 4.436,
    "args": {
      "External id": 857,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 434,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 434, "pid": 49625, "tid": 49625, "ts": 7253740879945.459,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879953.967, "dur": 3.922,
    "args": {
      "External id": 858,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 435,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 435, "pid": 49625, "tid": 49625, "ts": 7253740879953.967,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879961.307, "dur": 3.491,
    "args": {
      "External id": 859,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 436,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 436, "pid": 49625, "tid": 49625, "ts": 7253740879961.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879968.951, "dur": 4.011,
    "args": {
      "External id": 860,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 437,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 437, "pid": 49625, "tid": 49625, "ts": 7253740879968.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879977.760, "dur": 3.430,
    "args": {
      "External id": 861,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 438,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 438, "pid": 49625, "tid": 49625, "ts": 7253740879977.760,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879985.077, "dur": 3.211,
    "args": {
      "External id": 863,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 439,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 439, "pid": 49625, "tid": 49625, "ts": 7253740879985.077,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740879991.530, "dur": 4.026,
    "args": {
      "External id": 864,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 440,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 440, "pid": 49625, "tid": 49625, "ts": 7253740879991.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740880005.102, "dur": 4.215,
    "args": {
      "External id": 867,      "cid": 134, "correlation": 441, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 441, "pid": 49625, "tid": 49625, "ts": 7253740880005.102,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880032.730, "dur": 3.748,
    "args": {
      "External id": 874,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 442,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 442, "pid": 49625, "tid": 49625, "ts": 7253740880032.730,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880040.862, "dur": 3.786,
    "args": {
      "External id": 875,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 443,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 443, "pid": 49625, "tid": 49625, "ts": 7253740880040.862,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880048.772, "dur": 3.494,
    "args": {
      "External id": 876,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 444,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 444, "pid": 49625, "tid": 49625, "ts": 7253740880048.772,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880055.678, "dur": 3.350,
    "args": {
      "External id": 877,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 445,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 445, "pid": 49625, "tid": 49625, "ts": 7253740880055.678,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880063.071, "dur": 3.739,
    "args": {
      "External id": 878,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 446,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 446, "pid": 49625, "tid": 49625, "ts": 7253740880063.071,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880071.722, "dur": 3.169,
    "args": {
      "External id": 879,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 447,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 447, "pid": 49625, "tid": 49625, "ts": 7253740880071.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880078.809, "dur": 4.151,
    "args": {
      "External id": 881,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 448,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 448, "pid": 49625, "tid": 49625, "ts": 7253740880078.809,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880086.242, "dur": 4.007,
    "args": {
      "External id": 882,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 449,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 449, "pid": 49625, "tid": 49625, "ts": 7253740880086.242,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740880100.040, "dur": 4.201,
    "args": {
      "External id": 885,      "cid": 134, "correlation": 450, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 450, "pid": 49625, "tid": 49625, "ts": 7253740880100.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880127.942, "dur": 3.836,
    "args": {
      "External id": 892,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 451,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 451, "pid": 49625, "tid": 49625, "ts": 7253740880127.942,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880136.294, "dur": 3.314,
    "args": {
      "External id": 893,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 452,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 452, "pid": 49625, "tid": 49625, "ts": 7253740880136.294,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880143.680, "dur": 3.653,
    "args": {
      "External id": 894,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 453,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 453, "pid": 49625, "tid": 49625, "ts": 7253740880143.680,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880151.043, "dur": 3.844,
    "args": {
      "External id": 895,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 454,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 454, "pid": 49625, "tid": 49625, "ts": 7253740880151.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880158.958, "dur": 4.389,
    "args": {
      "External id": 896,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 455,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 455, "pid": 49625, "tid": 49625, "ts": 7253740880158.958,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880168.164, "dur": 3.757,
    "args": {
      "External id": 897,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 456,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 456, "pid": 49625, "tid": 49625, "ts": 7253740880168.164,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880175.838, "dur": 3.672,
    "args": {
      "External id": 899,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 457,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 457, "pid": 49625, "tid": 49625, "ts": 7253740880175.838,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740880183.787, "dur": 3.904,
    "args": {
      "External id": 900,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 458,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 458, "pid": 49625, "tid": 49625, "ts": 7253740880183.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740880298.741, "dur": 20573.110,
    "args": {
            "cid": 48, "correlation": 459
    }
  },
  {
    "ph": "f", "id": 459, "pid": 49625, "tid": 49625, "ts": 7253740880298.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740837093.835, "dur": 63.064,
    "args": {
      "External id": 3,      "device": 2, "stream": 0,      "correlation": 3, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2128.2443193530485
    }
  },
  {
    "ph": "f", "id": 3, "pid": 2, "tid": 0, "ts": 7253740837093.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740845015.148, "dur": 3.006,
    "args": {
      "External id": 10,      "device": 2, "stream": 0,      "correlation": 4, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 4, "pid": 2, "tid": 0, "ts": 7253740845015.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740850792.798, "dur": 2.846,
    "args": {
      "External id": 11,      "device": 2, "stream": 0,      "correlation": 5, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 5, "pid": 2, "tid": 0, "ts": 7253740850792.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740850929.512, "dur": 84.954,
    "args": {
      "External id": 12,      "device": 2, "stream": 0,      "correlation": 8, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 8, "pid": 2, "tid": 0, "ts": 7253740850929.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740851014.467, "dur": 64.949,
    "args": {
      "External id": 13,      "device": 2, "stream": 0,      "correlation": 11, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 11, "pid": 2, "tid": 0, "ts": 7253740851014.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740861146.719, "dur": 81.065,
    "args": {
      "External id": 14,      "device": 2, "stream": 0,      "correlation": 12, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 12, "pid": 2, "tid": 0, "ts": 7253740861146.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740874838.866, "dur": 63.785,
    "args": {
      "External id": 15,      "device": 2, "stream": 0,      "correlation": 13, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 13, "pid": 2, "tid": 0, "ts": 7253740874838.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740874906.742, "dur": 77.256,
    "args": {
      "External id": 17,      "device": 2, "stream": 0,      "correlation": 14, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 14, "pid": 2, "tid": 0, "ts": 7253740874906.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740874989.532, "dur": 75.613,
    "args": {
      "External id": 18,      "device": 2, "stream": 0,      "correlation": 17, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 17, "pid": 2, "tid": 0, "ts": 7253740874989.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740875065.146, "dur": 60.337,
    "args": {
      "External id": 21,      "device": 2, "stream": 0,      "correlation": 18, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2224.431171069641
    }
  },
  {
    "ph": "f", "id": 18, "pid": 2, "tid": 0, "ts": 7253740875065.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875171.390, "dur": 2.605,
    "args": {
      "External id": 28,      "device": 2, "stream": 0,      "correlation": 19, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 19, "pid": 2, "tid": 0, "ts": 7253740875171.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875188.990, "dur": 2.525,
    "args": {
      "External id": 29,      "device": 2, "stream": 0,      "correlation": 20, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 20, "pid": 2, "tid": 0, "ts": 7253740875188.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740875200.336, "dur": 78.139,
    "args": {
      "External id": 30,      "device": 2, "stream": 0,      "correlation": 21, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 21, "pid": 2, "tid": 0, "ts": 7253740875200.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875278.476, "dur": 63.785,
    "args": {
      "External id": 31,      "device": 2, "stream": 0,      "correlation": 22, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 22, "pid": 2, "tid": 0, "ts": 7253740875278.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740875342.262, "dur": 79.021,
    "args": {
      "External id": 32,      "device": 2, "stream": 0,      "correlation": 23, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 23, "pid": 2, "tid": 0, "ts": 7253740875342.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875421.284, "dur": 61.780,
    "args": {
      "External id": 33,      "device": 2, "stream": 0,      "correlation": 24, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 24, "pid": 2, "tid": 0, "ts": 7253740875421.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740901377.555, "dur": 2.627,
    "args": {
            "cid": 48, "correlation": 460
    }
  },
  {
    "ph": "f", "id": 460, "pid": 49625, "tid": 49625, "ts": 7253740901377.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740875483.065, "dur": 73.889,
    "args": {
      "External id": 35,      "device": 2, "stream": 0,      "correlation": 25, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 25, "pid": 2, "tid": 0, "ts": 7253740875483.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740875556.955, "dur": 75.853,
    "args": {
      "External id": 36,      "device": 2, "stream": 0,      "correlation": 26, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 26, "pid": 2, "tid": 0, "ts": 7253740875556.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740875632.809, "dur": 58.414,
    "args": {
      "External id": 39,      "device": 2, "stream": 0,      "correlation": 27, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2297.6586150817425
    }
  },
  {
    "ph": "f", "id": 27, "pid": 2, "tid": 0, "ts": 7253740875632.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875691.224, "dur": 5.571,
    "args": {
      "External id": 46,      "device": 2, "stream": 0,      "correlation": 28, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 28, "pid": 2, "tid": 0, "ts": 7253740875691.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875696.796, "dur": 5.452,
    "args": {
      "External id": 47,      "device": 2, "stream": 0,      "correlation": 29, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 29, "pid": 2, "tid": 0, "ts": 7253740875696.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740875702.249, "dur": 76.013,
    "args": {
      "External id": 48,      "device": 2, "stream": 0,      "correlation": 30, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 30, "pid": 2, "tid": 0, "ts": 7253740875702.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875778.263, "dur": 60.698,
    "args": {
      "External id": 49,      "device": 2, "stream": 0,      "correlation": 31, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 31, "pid": 2, "tid": 0, "ts": 7253740875778.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740875838.962, "dur": 75.372,
    "args": {
      "External id": 50,      "device": 2, "stream": 0,      "correlation": 32, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 32, "pid": 2, "tid": 0, "ts": 7253740875838.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740875914.335, "dur": 70.802,
    "args": {
      "External id": 51,      "device": 2, "stream": 0,      "correlation": 33, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 33, "pid": 2, "tid": 0, "ts": 7253740875914.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740875985.138, "dur": 74.730,
    "args": {
      "External id": 53,      "device": 2, "stream": 0,      "correlation": 34, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 34, "pid": 2, "tid": 0, "ts": 7253740875985.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740876059.869, "dur": 76.134,
    "args": {
      "External id": 54,      "device": 2, "stream": 0,      "correlation": 35, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 35, "pid": 2, "tid": 0, "ts": 7253740876059.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740876136.004, "dur": 61.941,
    "args": {
      "External id": 57,      "device": 2, "stream": 0,      "correlation": 36, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2166.829098188628
    }
  },
  {
    "ph": "f", "id": 36, "pid": 2, "tid": 0, "ts": 7253740876136.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876197.946, "dur": 5.612,
    "args": {
      "External id": 64,      "device": 2, "stream": 0,      "correlation": 37, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 37, "pid": 2, "tid": 0, "ts": 7253740876197.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876203.559, "dur": 5.492,
    "args": {
      "External id": 65,      "device": 2, "stream": 0,      "correlation": 38, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 38, "pid": 2, "tid": 0, "ts": 7253740876203.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740876209.052, "dur": 75.372,
    "args": {
      "External id": 66,      "device": 2, "stream": 0,      "correlation": 39, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 39, "pid": 2, "tid": 0, "ts": 7253740876209.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 49625, "tid": 49625,
    "ts": 7253740901455.164, "dur": 0.596,
    "args": {
      "External id": 901,      "cid": 290, "correlation": 461
    }
  },
  {
    "ph": "f", "id": 461, "pid": 49625, "tid": 49625, "ts": 7253740901455.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876284.425, "dur": 68.877,
    "args": {
      "External id": 67,      "device": 2, "stream": 0,      "correlation": 40, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 40, "pid": 2, "tid": 0, "ts": 7253740876284.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740876353.303, "dur": 76.134,
    "args": {
      "External id": 68,      "device": 2, "stream": 0,      "correlation": 41, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 41, "pid": 2, "tid": 0, "ts": 7253740876353.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876429.438, "dur": 65.750,
    "args": {
      "External id": 69,      "device": 2, "stream": 0,      "correlation": 42, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 42, "pid": 2, "tid": 0, "ts": 7253740876429.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740876495.189, "dur": 75.051,
    "args": {
      "External id": 71,      "device": 2, "stream": 0,      "correlation": 43, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 43, "pid": 2, "tid": 0, "ts": 7253740876495.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740876570.241, "dur": 74.330,
    "args": {
      "External id": 72,      "device": 2, "stream": 0,      "correlation": 44, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 44, "pid": 2, "tid": 0, "ts": 7253740876570.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740876644.572, "dur": 58.975,
    "args": {
      "External id": 75,      "device": 2, "stream": 0,      "correlation": 45, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2275.802495930548
    }
  },
  {
    "ph": "f", "id": 45, "pid": 2, "tid": 0, "ts": 7253740876644.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876703.548, "dur": 6.133,
    "args": {
      "External id": 82,      "device": 2, "stream": 0,      "correlation": 46, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 46, "pid": 2, "tid": 0, "ts": 7253740876703.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876709.682, "dur": 5.892,
    "args": {
      "External id": 83,      "device": 2, "stream": 0,      "correlation": 47, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 47, "pid": 2, "tid": 0, "ts": 7253740876709.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740876715.575, "dur": 76.575,
    "args": {
      "External id": 84,      "device": 2, "stream": 0,      "correlation": 48, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 48, "pid": 2, "tid": 0, "ts": 7253740876715.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876792.151, "dur": 69.719,
    "args": {
      "External id": 85,      "device": 2, "stream": 0,      "correlation": 49, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 49, "pid": 2, "tid": 0, "ts": 7253740876792.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740876861.871, "dur": 72.205,
    "args": {
      "External id": 86,      "device": 2, "stream": 0,      "correlation": 50, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 50, "pid": 2, "tid": 0, "ts": 7253740876861.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740876934.077, "dur": 67.234,
    "args": {
      "External id": 87,      "device": 2, "stream": 0,      "correlation": 51, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 51, "pid": 2, "tid": 0, "ts": 7253740876934.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740877001.312, "dur": 72.525,
    "args": {
      "External id": 89,      "device": 2, "stream": 0,      "correlation": 52, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 52, "pid": 2, "tid": 0, "ts": 7253740877001.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740877073.838, "dur": 75.613,
    "args": {
      "External id": 90,      "device": 2, "stream": 0,      "correlation": 53, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 53, "pid": 2, "tid": 0, "ts": 7253740877073.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740877149.452, "dur": 60.458,
    "args": {
      "External id": 93,      "device": 2, "stream": 0,      "correlation": 54, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2219.9792917514346
    }
  },
  {
    "ph": "f", "id": 54, "pid": 2, "tid": 0, "ts": 7253740877149.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877209.911, "dur": 5.772,
    "args": {
      "External id": 100,      "device": 2, "stream": 0,      "correlation": 55, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 55, "pid": 2, "tid": 0, "ts": 7253740877209.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877215.684, "dur": 5.572,
    "args": {
      "External id": 101,      "device": 2, "stream": 0,      "correlation": 56, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 56, "pid": 2, "tid": 0, "ts": 7253740877215.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740877221.257, "dur": 75.773,
    "args": {
      "External id": 102,      "device": 2, "stream": 0,      "correlation": 57, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 57, "pid": 2, "tid": 0, "ts": 7253740877221.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877297.031, "dur": 64.948,
    "args": {
      "External id": 103,      "device": 2, "stream": 0,      "correlation": 58, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 58, "pid": 2, "tid": 0, "ts": 7253740877297.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740877361.980, "dur": 72.285,
    "args": {
      "External id": 104,      "device": 2, "stream": 0,      "correlation": 59, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 59, "pid": 2, "tid": 0, "ts": 7253740877361.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 49625, "tid": 49625,
    "ts": 7253740901456.638, "dur": 165.583,
    "args": {
      "External id": 901,            "bytes": 268435456, "cid": 108, "correlation": 462, "ptr": "0x7f6dada21be8"
    }
  },
  {
    "ph": "s", "id": 462, "pid": 49625, "tid": 49625, "ts": 7253740901456.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877434.266, "dur": 63.585,
    "args": {
      "External id": 105,      "device": 2, "stream": 0,      "correlation": 60, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 60, "pid": 2, "tid": 0, "ts": 7253740877434.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740877497.852, "dur": 71.603,
    "args": {
      "External id": 107,      "device": 2, "stream": 0,      "correlation": 61, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 61, "pid": 2, "tid": 0, "ts": 7253740877497.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740877569.456, "dur": 76.214,
    "args": {
      "External id": 108,      "device": 2, "stream": 0,      "correlation": 62, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 62, "pid": 2, "tid": 0, "ts": 7253740877569.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740877645.671, "dur": 63.505,
    "args": {
      "External id": 111,      "device": 2, "stream": 0,      "correlation": 63, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2113.465310364375
    }
  },
  {
    "ph": "f", "id": 63, "pid": 2, "tid": 0, "ts": 7253740877645.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877709.177, "dur": 6.013,
    "args": {
      "External id": 118,      "device": 2, "stream": 0,      "correlation": 64, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 64, "pid": 2, "tid": 0, "ts": 7253740877709.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877715.191, "dur": 6.053,
    "args": {
      "External id": 119,      "device": 2, "stream": 0,      "correlation": 65, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 65, "pid": 2, "tid": 0, "ts": 7253740877715.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740877721.245, "dur": 81.667,
    "args": {
      "External id": 120,      "device": 2, "stream": 0,      "correlation": 66, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 66, "pid": 2, "tid": 0, "ts": 7253740877721.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877802.913, "dur": 61.099,
    "args": {
      "External id": 121,      "device": 2, "stream": 0,      "correlation": 67, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 67, "pid": 2, "tid": 0, "ts": 7253740877802.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740877864.013, "dur": 76.976,
    "args": {
      "External id": 122,      "device": 2, "stream": 0,      "correlation": 68, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 68, "pid": 2, "tid": 0, "ts": 7253740877864.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901659.853, "dur": 10.805,
    "args": {
      "External id": 903,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 463,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 463, "pid": 49625, "tid": 49625, "ts": 7253740901659.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740877940.990, "dur": 64.988,
    "args": {
      "External id": 123,      "device": 2, "stream": 0,      "correlation": 69, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 69, "pid": 2, "tid": 0, "ts": 7253740877940.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740878005.979, "dur": 76.254,
    "args": {
      "External id": 125,      "device": 2, "stream": 0,      "correlation": 70, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 70, "pid": 2, "tid": 0, "ts": 7253740878005.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740878082.234, "dur": 83.471,
    "args": {
      "External id": 126,      "device": 2, "stream": 0,      "correlation": 71, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 71, "pid": 2, "tid": 0, "ts": 7253740878082.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740878165.706, "dur": 60.859,
    "args": {
      "External id": 129,      "device": 2, "stream": 0,      "correlation": 72, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2205.3520867564903
    }
  },
  {
    "ph": "f", "id": 72, "pid": 2, "tid": 0, "ts": 7253740878165.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878226.566, "dur": 6.093,
    "args": {
      "External id": 136,      "device": 2, "stream": 0,      "correlation": 73, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 73, "pid": 2, "tid": 0, "ts": 7253740878226.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878232.660, "dur": 5.812,
    "args": {
      "External id": 137,      "device": 2, "stream": 0,      "correlation": 74, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 74, "pid": 2, "tid": 0, "ts": 7253740878232.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740901695.650, "dur": 8.651,
    "args": {
      "External id": 906,      "cid": 134, "correlation": 464, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 464, "pid": 49625, "tid": 49625, "ts": 7253740901695.650,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740878238.473, "dur": 83.711,
    "args": {
      "External id": 138,      "device": 2, "stream": 0,      "correlation": 75, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 75, "pid": 2, "tid": 0, "ts": 7253740878238.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878322.185, "dur": 63.906,
    "args": {
      "External id": 139,      "device": 2, "stream": 0,      "correlation": 76, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 76, "pid": 2, "tid": 0, "ts": 7253740878322.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740878386.092, "dur": 77.417,
    "args": {
      "External id": 140,      "device": 2, "stream": 0,      "correlation": 77, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 77, "pid": 2, "tid": 0, "ts": 7253740878386.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878463.510, "dur": 64.146,
    "args": {
      "External id": 141,      "device": 2, "stream": 0,      "correlation": 78, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 78, "pid": 2, "tid": 0, "ts": 7253740878463.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740878527.657, "dur": 76.575,
    "args": {
      "External id": 143,      "device": 2, "stream": 0,      "correlation": 79, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 79, "pid": 2, "tid": 0, "ts": 7253740878527.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740878604.233, "dur": 84.273,
    "args": {
      "External id": 144,      "device": 2, "stream": 0,      "correlation": 80, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 80, "pid": 2, "tid": 0, "ts": 7253740878604.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740878688.507, "dur": 60.738,
    "args": {
      "External id": 147,      "device": 2, "stream": 0,      "correlation": 81, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2209.7454353874778
    }
  },
  {
    "ph": "f", "id": 81, "pid": 2, "tid": 0, "ts": 7253740878688.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878749.246, "dur": 5.933,
    "args": {
      "External id": 154,      "device": 2, "stream": 0,      "correlation": 82, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 82, "pid": 2, "tid": 0, "ts": 7253740878749.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878755.180, "dur": 5.692,
    "args": {
      "External id": 155,      "device": 2, "stream": 0,      "correlation": 83, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 83, "pid": 2, "tid": 0, "ts": 7253740878755.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740878760.873, "dur": 87.881,
    "args": {
      "External id": 156,      "device": 2, "stream": 0,      "correlation": 84, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 84, "pid": 2, "tid": 0, "ts": 7253740878760.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878848.755, "dur": 60.618,
    "args": {
      "External id": 157,      "device": 2, "stream": 0,      "correlation": 85, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 85, "pid": 2, "tid": 0, "ts": 7253740878848.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740878909.374, "dur": 81.707,
    "args": {
      "External id": 158,      "device": 2, "stream": 0,      "correlation": 86, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 86, "pid": 2, "tid": 0, "ts": 7253740878909.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740878991.082, "dur": 65.870,
    "args": {
      "External id": 159,      "device": 2, "stream": 0,      "correlation": 87, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 87, "pid": 2, "tid": 0, "ts": 7253740878991.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740879056.953, "dur": 80.544,
    "args": {
      "External id": 161,      "device": 2, "stream": 0,      "correlation": 88, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 88, "pid": 2, "tid": 0, "ts": 7253740879056.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740879137.498, "dur": 90.406,
    "args": {
      "External id": 162,      "device": 2, "stream": 0,      "correlation": 89, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 89, "pid": 2, "tid": 0, "ts": 7253740879137.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740879227.905, "dur": 62.142,
    "args": {
      "External id": 165,      "device": 2, "stream": 0,      "correlation": 90, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2159.8205429412806
    }
  },
  {
    "ph": "f", "id": 90, "pid": 2, "tid": 0, "ts": 7253740879227.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879290.048, "dur": 5.813,
    "args": {
      "External id": 172,      "device": 2, "stream": 0,      "correlation": 91, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 91, "pid": 2, "tid": 0, "ts": 7253740879290.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879295.862, "dur": 5.451,
    "args": {
      "External id": 173,      "device": 2, "stream": 0,      "correlation": 92, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 92, "pid": 2, "tid": 0, "ts": 7253740879295.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740879301.314, "dur": 90.326,
    "args": {
      "External id": 174,      "device": 2, "stream": 0,      "correlation": 93, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 93, "pid": 2, "tid": 0, "ts": 7253740879301.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879391.641, "dur": 61.421,
    "args": {
      "External id": 175,      "device": 2, "stream": 0,      "correlation": 94, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 94, "pid": 2, "tid": 0, "ts": 7253740879391.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740879453.063, "dur": 81.786,
    "args": {
      "External id": 176,      "device": 2, "stream": 0,      "correlation": 95, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 95, "pid": 2, "tid": 0, "ts": 7253740879453.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901851.023, "dur": 7.848,
    "args": {
      "External id": 913,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 465,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 465, "pid": 49625, "tid": 49625, "ts": 7253740901851.023,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879534.850, "dur": 64.828,
    "args": {
      "External id": 177,      "device": 2, "stream": 0,      "correlation": 96, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 96, "pid": 2, "tid": 0, "ts": 7253740879534.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740879599.679, "dur": 81.466,
    "args": {
      "External id": 179,      "device": 2, "stream": 0,      "correlation": 97, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 97, "pid": 2, "tid": 0, "ts": 7253740879599.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740879681.146, "dur": 92.171,
    "args": {
      "External id": 180,      "device": 2, "stream": 0,      "correlation": 98, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 98, "pid": 2, "tid": 0, "ts": 7253740879681.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740879773.318, "dur": 60.658,
    "args": {
      "External id": 183,      "device": 2, "stream": 0,      "correlation": 99, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2212.6597537051384
    }
  },
  {
    "ph": "f", "id": 99, "pid": 2, "tid": 0, "ts": 7253740879773.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879833.977, "dur": 6.013,
    "args": {
      "External id": 190,      "device": 2, "stream": 0,      "correlation": 100, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 100, "pid": 2, "tid": 0, "ts": 7253740879833.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901876.908, "dur": 8.785,
    "args": {
      "External id": 914,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 466,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 466, "pid": 49625, "tid": 49625, "ts": 7253740901876.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879839.991, "dur": 5.612,
    "args": {
      "External id": 191,      "device": 2, "stream": 0,      "correlation": 101, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 101, "pid": 2, "tid": 0, "ts": 7253740879839.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740879845.604, "dur": 93.293,
    "args": {
      "External id": 192,      "device": 2, "stream": 0,      "correlation": 102, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 102, "pid": 2, "tid": 0, "ts": 7253740879845.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901894.697, "dur": 5.627,
    "args": {
      "External id": 915,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 467,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 467, "pid": 49625, "tid": 49625, "ts": 7253740901894.697,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740879938.898, "dur": 58.293,
    "args": {
      "External id": 193,      "device": 2, "stream": 0,      "correlation": 103, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 103, "pid": 2, "tid": 0, "ts": 7253740879938.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740879997.192, "dur": 83.230,
    "args": {
      "External id": 194,      "device": 2, "stream": 0,      "correlation": 104, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 104, "pid": 2, "tid": 0, "ts": 7253740879997.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901908.077, "dur": 4.891,
    "args": {
      "External id": 916,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 468,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 468, "pid": 49625, "tid": 49625, "ts": 7253740901908.077,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880080.423, "dur": 63.305,
    "args": {
      "External id": 195,      "device": 2, "stream": 0,      "correlation": 105, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 105, "pid": 2, "tid": 0, "ts": 7253740880080.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740880143.729, "dur": 83.952,
    "args": {
      "External id": 197,      "device": 2, "stream": 0,      "correlation": 106, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 106, "pid": 2, "tid": 0, "ts": 7253740880143.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740880227.682, "dur": 92.892,
    "args": {
      "External id": 198,      "device": 2, "stream": 0,      "correlation": 107, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 107, "pid": 2, "tid": 0, "ts": 7253740880227.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740880320.575, "dur": 62.703,
    "args": {
      "External id": 201,      "device": 2, "stream": 0,      "correlation": 108, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2140.4970655779534
    }
  },
  {
    "ph": "f", "id": 108, "pid": 2, "tid": 0, "ts": 7253740880320.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901922.720, "dur": 5.226,
    "args": {
      "External id": 917,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 469,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 469, "pid": 49625, "tid": 49625, "ts": 7253740901922.720,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880383.279, "dur": 6.133,
    "args": {
      "External id": 208,      "device": 2, "stream": 0,      "correlation": 109, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 109, "pid": 2, "tid": 0, "ts": 7253740880383.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880389.413, "dur": 5.973,
    "args": {
      "External id": 209,      "device": 2, "stream": 0,      "correlation": 110, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 110, "pid": 2, "tid": 0, "ts": 7253740880389.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740880395.387, "dur": 93.453,
    "args": {
      "External id": 210,      "device": 2, "stream": 0,      "correlation": 111, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 111, "pid": 2, "tid": 0, "ts": 7253740880395.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901940.029, "dur": 5.364,
    "args": {
      "External id": 918,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 470,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 470, "pid": 49625, "tid": 49625, "ts": 7253740901940.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880488.841, "dur": 64.748,
    "args": {
      "External id": 211,      "device": 2, "stream": 0,      "correlation": 112, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 112, "pid": 2, "tid": 0, "ts": 7253740880488.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740880553.590, "dur": 83.671,
    "args": {
      "External id": 212,      "device": 2, "stream": 0,      "correlation": 113, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 113, "pid": 2, "tid": 0, "ts": 7253740880553.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901953.542, "dur": 3.959,
    "args": {
      "External id": 920,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 471,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 471, "pid": 49625, "tid": 49625, "ts": 7253740901953.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880637.262, "dur": 66.071,
    "args": {
      "External id": 213,      "device": 2, "stream": 0,      "correlation": 114, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 114, "pid": 2, "tid": 0, "ts": 7253740880637.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740901961.995, "dur": 4.115,
    "args": {
      "External id": 921,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 472,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 472, "pid": 49625, "tid": 49625, "ts": 7253740901961.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740880703.334, "dur": 83.070,
    "args": {
      "External id": 215,      "device": 2, "stream": 0,      "correlation": 115, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 115, "pid": 2, "tid": 0, "ts": 7253740880703.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740880786.405, "dur": 93.373,
    "args": {
      "External id": 216,      "device": 2, "stream": 0,      "correlation": 116, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 116, "pid": 2, "tid": 0, "ts": 7253740880786.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740880879.779, "dur": 60.779,
    "args": {
      "External id": 219,      "device": 2, "stream": 0,      "correlation": 117, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2208.2548206646925
    }
  },
  {
    "ph": "f", "id": 117, "pid": 2, "tid": 0, "ts": 7253740880879.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880940.559, "dur": 5.892,
    "args": {
      "External id": 226,      "device": 2, "stream": 0,      "correlation": 118, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 118, "pid": 2, "tid": 0, "ts": 7253740880940.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740880946.452, "dur": 5.893,
    "args": {
      "External id": 227,      "device": 2, "stream": 0,      "correlation": 119, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 119, "pid": 2, "tid": 0, "ts": 7253740880946.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740880952.346, "dur": 93.454,
    "args": {
      "External id": 228,      "device": 2, "stream": 0,      "correlation": 120, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 120, "pid": 2, "tid": 0, "ts": 7253740880952.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881045.801, "dur": 62.943,
    "args": {
      "External id": 229,      "device": 2, "stream": 0,      "correlation": 121, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 121, "pid": 2, "tid": 0, "ts": 7253740881045.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740881108.745, "dur": 84.513,
    "args": {
      "External id": 230,      "device": 2, "stream": 0,      "correlation": 122, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 122, "pid": 2, "tid": 0, "ts": 7253740881108.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881193.259, "dur": 65.630,
    "args": {
      "External id": 231,      "device": 2, "stream": 0,      "correlation": 123, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 123, "pid": 2, "tid": 0, "ts": 7253740881193.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902010.800, "dur": 4.928,
    "args": {
      "External id": 923,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 473,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 473, "pid": 49625, "tid": 49625, "ts": 7253740902010.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740881258.890, "dur": 83.310,
    "args": {
      "External id": 233,      "device": 2, "stream": 0,      "correlation": 124, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 124, "pid": 2, "tid": 0, "ts": 7253740881258.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740881342.201, "dur": 93.053,
    "args": {
      "External id": 234,      "device": 2, "stream": 0,      "correlation": 125, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 125, "pid": 2, "tid": 0, "ts": 7253740881342.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740881435.255, "dur": 62.783,
    "args": {
      "External id": 237,      "device": 2, "stream": 0,      "correlation": 126, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2137.769622833843
    }
  },
  {
    "ph": "f", "id": 126, "pid": 2, "tid": 0, "ts": 7253740881435.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881498.039, "dur": 5.733,
    "args": {
      "External id": 244,      "device": 2, "stream": 0,      "correlation": 127, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 127, "pid": 2, "tid": 0, "ts": 7253740881498.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881503.773, "dur": 6.012,
    "args": {
      "External id": 245,      "device": 2, "stream": 0,      "correlation": 128, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 128, "pid": 2, "tid": 0, "ts": 7253740881503.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902029.821, "dur": 5.514,
    "args": {
      "External id": 926,      "cid": 134, "correlation": 474, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 474, "pid": 49625, "tid": 49625, "ts": 7253740902029.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740881509.786, "dur": 93.534,
    "args": {
      "External id": 246,      "device": 2, "stream": 0,      "correlation": 129, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 129, "pid": 2, "tid": 0, "ts": 7253740881509.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881603.321, "dur": 67.514,
    "args": {
      "External id": 247,      "device": 2, "stream": 0,      "correlation": 130, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 130, "pid": 2, "tid": 0, "ts": 7253740881603.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740881670.836, "dur": 84.874,
    "args": {
      "External id": 248,      "device": 2, "stream": 0,      "correlation": 131, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 131, "pid": 2, "tid": 0, "ts": 7253740881670.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740881755.711, "dur": 62.222,
    "args": {
      "External id": 249,      "device": 2, "stream": 0,      "correlation": 132, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 132, "pid": 2, "tid": 0, "ts": 7253740881755.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740881817.934, "dur": 81.386,
    "args": {
      "External id": 251,      "device": 2, "stream": 0,      "correlation": 133, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 133, "pid": 2, "tid": 0, "ts": 7253740881817.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740881899.321, "dur": 90.447,
    "args": {
      "External id": 252,      "device": 2, "stream": 0,      "correlation": 134, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 134, "pid": 2, "tid": 0, "ts": 7253740881899.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740881989.769, "dur": 57.972,
    "args": {
      "External id": 255,      "device": 2, "stream": 0,      "correlation": 135, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2315.176513204423
    }
  },
  {
    "ph": "f", "id": 135, "pid": 2, "tid": 0, "ts": 7253740881989.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882047.742, "dur": 5.892,
    "args": {
      "External id": 262,      "device": 2, "stream": 0,      "correlation": 136, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 136, "pid": 2, "tid": 0, "ts": 7253740882047.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882053.635, "dur": 5.572,
    "args": {
      "External id": 263,      "device": 2, "stream": 0,      "correlation": 137, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 137, "pid": 2, "tid": 0, "ts": 7253740882053.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740882059.208, "dur": 90.808,
    "args": {
      "External id": 264,      "device": 2, "stream": 0,      "correlation": 138, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 138, "pid": 2, "tid": 0, "ts": 7253740882059.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882150.017, "dur": 59.736,
    "args": {
      "External id": 265,      "device": 2, "stream": 0,      "correlation": 139, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 139, "pid": 2, "tid": 0, "ts": 7253740882150.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740882209.754, "dur": 83.511,
    "args": {
      "External id": 266,      "device": 2, "stream": 0,      "correlation": 140, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 140, "pid": 2, "tid": 0, "ts": 7253740882209.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882293.266, "dur": 65.549,
    "args": {
      "External id": 267,      "device": 2, "stream": 0,      "correlation": 141, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 141, "pid": 2, "tid": 0, "ts": 7253740882293.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740882358.816, "dur": 81.627,
    "args": {
      "External id": 269,      "device": 2, "stream": 0,      "correlation": 142, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 142, "pid": 2, "tid": 0, "ts": 7253740882358.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902124.273, "dur": 5.734,
    "args": {
      "External id": 933,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 475,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 475, "pid": 49625, "tid": 49625, "ts": 7253740902124.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740882440.444, "dur": 91.048,
    "args": {
      "External id": 270,      "device": 2, "stream": 0,      "correlation": 143, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 143, "pid": 2, "tid": 0, "ts": 7253740882440.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740882531.493, "dur": 58.854,
    "args": {
      "External id": 273,      "device": 2, "stream": 0,      "correlation": 144, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2280.4813184946056
    }
  },
  {
    "ph": "f", "id": 144, "pid": 2, "tid": 0, "ts": 7253740882531.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882590.348, "dur": 6.133,
    "args": {
      "External id": 280,      "device": 2, "stream": 0,      "correlation": 145, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 145, "pid": 2, "tid": 0, "ts": 7253740882590.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902139.144, "dur": 4.644,
    "args": {
      "External id": 934,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 476,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 476, "pid": 49625, "tid": 49625, "ts": 7253740902139.144,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882596.482, "dur": 6.013,
    "args": {
      "External id": 281,      "device": 2, "stream": 0,      "correlation": 146, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 146, "pid": 2, "tid": 0, "ts": 7253740882596.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740882602.496, "dur": 90.888,
    "args": {
      "External id": 282,      "device": 2, "stream": 0,      "correlation": 147, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 147, "pid": 2, "tid": 0, "ts": 7253740882602.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902150.579, "dur": 4.418,
    "args": {
      "External id": 935,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 477,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 477, "pid": 49625, "tid": 49625, "ts": 7253740902150.579,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882693.385, "dur": 62.222,
    "args": {
      "External id": 283,      "device": 2, "stream": 0,      "correlation": 148, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 148, "pid": 2, "tid": 0, "ts": 7253740882693.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902159.599, "dur": 3.821,
    "args": {
      "External id": 936,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 478,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 478, "pid": 49625, "tid": 49625, "ts": 7253740902159.599,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740882755.608, "dur": 83.230,
    "args": {
      "External id": 284,      "device": 2, "stream": 0,      "correlation": 149, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 149, "pid": 2, "tid": 0, "ts": 7253740882755.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740882838.839, "dur": 64.948,
    "args": {
      "External id": 285,      "device": 2, "stream": 0,      "correlation": 150, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 150, "pid": 2, "tid": 0, "ts": 7253740882838.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902169.211, "dur": 4.158,
    "args": {
      "External id": 937,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 479,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 479, "pid": 49625, "tid": 49625, "ts": 7253740902169.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740882903.788, "dur": 80.704,
    "args": {
      "External id": 287,      "device": 2, "stream": 0,      "correlation": 151, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 151, "pid": 2, "tid": 0, "ts": 7253740882903.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740882984.493, "dur": 90.808,
    "args": {
      "External id": 288,      "device": 2, "stream": 0,      "correlation": 152, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 152, "pid": 2, "tid": 0, "ts": 7253740882984.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740883075.302, "dur": 60.137,
    "args": {
      "External id": 291,      "device": 2, "stream": 0,      "correlation": 153, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2231.8289268016892
    }
  },
  {
    "ph": "f", "id": 153, "pid": 2, "tid": 0, "ts": 7253740883075.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902180.607, "dur": 3.489,
    "args": {
      "External id": 938,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 480,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 480, "pid": 49625, "tid": 49625, "ts": 7253740902180.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883135.440, "dur": 6.133,
    "args": {
      "External id": 298,      "device": 2, "stream": 0,      "correlation": 154, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 154, "pid": 2, "tid": 0, "ts": 7253740883135.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902189.369, "dur": 3.668,
    "args": {
      "External id": 940,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 481,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 481, "pid": 49625, "tid": 49625, "ts": 7253740902189.369,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883141.574, "dur": 5.772,
    "args": {
      "External id": 299,      "device": 2, "stream": 0,      "correlation": 155, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 155, "pid": 2, "tid": 0, "ts": 7253740883141.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902197.540, "dur": 3.959,
    "args": {
      "External id": 941,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 482,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 482, "pid": 49625, "tid": 49625, "ts": 7253740902197.540,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740883147.347, "dur": 90.247,
    "args": {
      "External id": 300,      "device": 2, "stream": 0,      "correlation": 156, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 156, "pid": 2, "tid": 0, "ts": 7253740883147.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883237.595, "dur": 59.936,
    "args": {
      "External id": 301,      "device": 2, "stream": 0,      "correlation": 157, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 157, "pid": 2, "tid": 0, "ts": 7253740883237.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740883297.532, "dur": 81.707,
    "args": {
      "External id": 302,      "device": 2, "stream": 0,      "correlation": 158, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 158, "pid": 2, "tid": 0, "ts": 7253740883297.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883379.240, "dur": 65.429,
    "args": {
      "External id": 303,      "device": 2, "stream": 0,      "correlation": 159, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 159, "pid": 2, "tid": 0, "ts": 7253740883379.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740883444.670, "dur": 80.825,
    "args": {
      "External id": 305,      "device": 2, "stream": 0,      "correlation": 160, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 160, "pid": 2, "tid": 0, "ts": 7253740883444.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740883525.496, "dur": 90.286,
    "args": {
      "External id": 306,      "device": 2, "stream": 0,      "correlation": 161, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 161, "pid": 2, "tid": 0, "ts": 7253740883525.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740883615.783, "dur": 63.144,
    "args": {
      "External id": 309,      "device": 2, "stream": 0,      "correlation": 162, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2125.547992715179
    }
  },
  {
    "ph": "f", "id": 162, "pid": 2, "tid": 0, "ts": 7253740883615.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883678.928, "dur": 5.933,
    "args": {
      "External id": 316,      "device": 2, "stream": 0,      "correlation": 163, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 163, "pid": 2, "tid": 0, "ts": 7253740883678.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902236.712, "dur": 4.203,
    "args": {
      "External id": 943,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 483,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 483, "pid": 49625, "tid": 49625, "ts": 7253740902236.712,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883684.862, "dur": 5.612,
    "args": {
      "External id": 317,      "device": 2, "stream": 0,      "correlation": 164, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 164, "pid": 2, "tid": 0, "ts": 7253740883684.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902251.622, "dur": 4.366,
    "args": {
      "External id": 946,      "cid": 134, "correlation": 484, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 484, "pid": 49625, "tid": 49625, "ts": 7253740902251.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740883690.475, "dur": 90.567,
    "args": {
      "External id": 318,      "device": 2, "stream": 0,      "correlation": 165, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 165, "pid": 2, "tid": 0, "ts": 7253740883690.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883781.043, "dur": 59.215,
    "args": {
      "External id": 319,      "device": 2, "stream": 0,      "correlation": 166, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 166, "pid": 2, "tid": 0, "ts": 7253740883781.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740883840.259, "dur": 81.426,
    "args": {
      "External id": 320,      "device": 2, "stream": 0,      "correlation": 167, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 167, "pid": 2, "tid": 0, "ts": 7253740883840.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740883921.686, "dur": 63.946,
    "args": {
      "External id": 321,      "device": 2, "stream": 0,      "correlation": 168, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 168, "pid": 2, "tid": 0, "ts": 7253740883921.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902313.889, "dur": 4.544,
    "args": {
      "External id": 953,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 485,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 485, "pid": 49625, "tid": 49625, "ts": 7253740902313.889,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740883985.633, "dur": 81.546,
    "args": {
      "External id": 323,      "device": 2, "stream": 0,      "correlation": 169, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 169, "pid": 2, "tid": 0, "ts": 7253740883985.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740884067.180, "dur": 91.249,
    "args": {
      "External id": 324,      "device": 2, "stream": 0,      "correlation": 170, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 170, "pid": 2, "tid": 0, "ts": 7253740884067.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740884158.430, "dur": 60.618,
    "args": {
      "External id": 327,      "device": 2, "stream": 0,      "correlation": 171, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2214.11979742325
    }
  },
  {
    "ph": "f", "id": 171, "pid": 2, "tid": 0, "ts": 7253740884158.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902327.137, "dur": 4.976,
    "args": {
      "External id": 954,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 486,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 486, "pid": 49625, "tid": 49625, "ts": 7253740902327.137,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884219.049, "dur": 6.133,
    "args": {
      "External id": 334,      "device": 2, "stream": 0,      "correlation": 172, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 172, "pid": 2, "tid": 0, "ts": 7253740884219.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884225.183, "dur": 5.451,
    "args": {
      "External id": 335,      "device": 2, "stream": 0,      "correlation": 173, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 173, "pid": 2, "tid": 0, "ts": 7253740884225.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902338.962, "dur": 4.289,
    "args": {
      "External id": 955,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 487,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 487, "pid": 49625, "tid": 49625, "ts": 7253740902338.962,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740884230.635, "dur": 90.688,
    "args": {
      "External id": 336,      "device": 2, "stream": 0,      "correlation": 174, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 174, "pid": 2, "tid": 0, "ts": 7253740884230.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902347.549, "dur": 3.561,
    "args": {
      "External id": 956,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 488,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 488, "pid": 49625, "tid": 49625, "ts": 7253740902347.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884321.324, "dur": 59.215,
    "args": {
      "External id": 337,      "device": 2, "stream": 0,      "correlation": 175, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 175, "pid": 2, "tid": 0, "ts": 7253740884321.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740884380.540, "dur": 81.626,
    "args": {
      "External id": 338,      "device": 2, "stream": 0,      "correlation": 176, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 176, "pid": 2, "tid": 0, "ts": 7253740884380.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902356.551, "dur": 3.824,
    "args": {
      "External id": 957,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 489,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 489, "pid": 49625, "tid": 49625, "ts": 7253740902356.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884462.167, "dur": 65.991,
    "args": {
      "External id": 339,      "device": 2, "stream": 0,      "correlation": 177, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 177, "pid": 2, "tid": 0, "ts": 7253740884462.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902367.446, "dur": 3.669,
    "args": {
      "External id": 958,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 490,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 490, "pid": 49625, "tid": 49625, "ts": 7253740902367.446,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740884528.159, "dur": 81.105,
    "args": {
      "External id": 341,      "device": 2, "stream": 0,      "correlation": 178, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 178, "pid": 2, "tid": 0, "ts": 7253740884528.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902376.324, "dur": 3.934,
    "args": {
      "External id": 960,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 491,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 491, "pid": 49625, "tid": 49625, "ts": 7253740902376.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740884609.265, "dur": 90.326,
    "args": {
      "External id": 342,      "device": 2, "stream": 0,      "correlation": 179, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 179, "pid": 2, "tid": 0, "ts": 7253740884609.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740884699.592, "dur": 61.421,
    "args": {
      "External id": 345,      "device": 2, "stream": 0,      "correlation": 180, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2185.173520888281
    }
  },
  {
    "ph": "f", "id": 180, "pid": 2, "tid": 0, "ts": 7253740884699.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884761.014, "dur": 6.052,
    "args": {
      "External id": 352,      "device": 2, "stream": 0,      "correlation": 181, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 181, "pid": 2, "tid": 0, "ts": 7253740884761.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902384.862, "dur": 4.145,
    "args": {
      "External id": 961,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 492,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 492, "pid": 49625, "tid": 49625, "ts": 7253740902384.862,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884767.067, "dur": 5.853,
    "args": {
      "External id": 353,      "device": 2, "stream": 0,      "correlation": 182, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 182, "pid": 2, "tid": 0, "ts": 7253740884767.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740884772.921, "dur": 89.444,
    "args": {
      "External id": 354,      "device": 2, "stream": 0,      "correlation": 183, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 183, "pid": 2, "tid": 0, "ts": 7253740884772.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740884862.366, "dur": 60.017,
    "args": {
      "External id": 355,      "device": 2, "stream": 0,      "correlation": 184, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 184, "pid": 2, "tid": 0, "ts": 7253740884862.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740884922.384, "dur": 80.304,
    "args": {
      "External id": 356,      "device": 2, "stream": 0,      "correlation": 185, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 185, "pid": 2, "tid": 0, "ts": 7253740884922.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885002.689, "dur": 65.389,
    "args": {
      "External id": 357,      "device": 2, "stream": 0,      "correlation": 186, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 186, "pid": 2, "tid": 0, "ts": 7253740885002.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902418.032, "dur": 3.632,
    "args": {
      "External id": 963,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 493,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 493, "pid": 49625, "tid": 49625, "ts": 7253740902418.032,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740885068.079, "dur": 80.584,
    "args": {
      "External id": 359,      "device": 2, "stream": 0,      "correlation": 187, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 187, "pid": 2, "tid": 0, "ts": 7253740885068.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740885148.664, "dur": 88.963,
    "args": {
      "External id": 360,      "device": 2, "stream": 0,      "correlation": 188, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 188, "pid": 2, "tid": 0, "ts": 7253740885148.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740885237.628, "dur": 60.739,
    "args": {
      "External id": 363,      "device": 2, "stream": 0,      "correlation": 189, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2209.7090549884756
    }
  },
  {
    "ph": "f", "id": 189, "pid": 2, "tid": 0, "ts": 7253740885237.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902431.147, "dur": 3.965,
    "args": {
      "External id": 966,      "cid": 134, "correlation": 494, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 494, "pid": 49625, "tid": 49625, "ts": 7253740902431.147,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885298.368, "dur": 6.093,
    "args": {
      "External id": 370,      "device": 2, "stream": 0,      "correlation": 190, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 190, "pid": 2, "tid": 0, "ts": 7253740885298.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885304.462, "dur": 5.692,
    "args": {
      "External id": 371,      "device": 2, "stream": 0,      "correlation": 191, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 191, "pid": 2, "tid": 0, "ts": 7253740885304.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740885310.155, "dur": 88.763,
    "args": {
      "External id": 372,      "device": 2, "stream": 0,      "correlation": 192, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 192, "pid": 2, "tid": 0, "ts": 7253740885310.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885398.919, "dur": 60.778,
    "args": {
      "External id": 373,      "device": 2, "stream": 0,      "correlation": 193, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 193, "pid": 2, "tid": 0, "ts": 7253740885398.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740885459.698, "dur": 80.103,
    "args": {
      "External id": 374,      "device": 2, "stream": 0,      "correlation": 194, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 194, "pid": 2, "tid": 0, "ts": 7253740885459.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885539.802, "dur": 66.672,
    "args": {
      "External id": 375,      "device": 2, "stream": 0,      "correlation": 195, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 195, "pid": 2, "tid": 0, "ts": 7253740885539.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740885606.475, "dur": 79.983,
    "args": {
      "External id": 377,      "device": 2, "stream": 0,      "correlation": 196, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 196, "pid": 2, "tid": 0, "ts": 7253740885606.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740885686.459, "dur": 88.242,
    "args": {
      "External id": 378,      "device": 2, "stream": 0,      "correlation": 197, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 197, "pid": 2, "tid": 0, "ts": 7253740885686.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740885774.702, "dur": 61.660,
    "args": {
      "External id": 381,      "device": 2, "stream": 0,      "correlation": 198, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2176.7037187200985
    }
  },
  {
    "ph": "f", "id": 198, "pid": 2, "tid": 0, "ts": 7253740885774.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902481.973, "dur": 4.246,
    "args": {
      "External id": 973,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 495,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 495, "pid": 49625, "tid": 49625, "ts": 7253740902481.973,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885836.363, "dur": 5.813,
    "args": {
      "External id": 388,      "device": 2, "stream": 0,      "correlation": 199, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 199, "pid": 2, "tid": 0, "ts": 7253740885836.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885842.177, "dur": 5.572,
    "args": {
      "External id": 389,      "device": 2, "stream": 0,      "correlation": 200, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 200, "pid": 2, "tid": 0, "ts": 7253740885842.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902491.789, "dur": 3.947,
    "args": {
      "External id": 974,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 496,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 496, "pid": 49625, "tid": 49625, "ts": 7253740902491.789,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740885847.750, "dur": 87.800,
    "args": {
      "External id": 390,      "device": 2, "stream": 0,      "correlation": 201, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 201, "pid": 2, "tid": 0, "ts": 7253740885847.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740885935.551, "dur": 65.470,
    "args": {
      "External id": 391,      "device": 2, "stream": 0,      "correlation": 202, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 202, "pid": 2, "tid": 0, "ts": 7253740885935.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902500.710, "dur": 3.220,
    "args": {
      "External id": 975,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 497,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 497, "pid": 49625, "tid": 49625, "ts": 7253740902500.710,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740886001.022, "dur": 80.223,
    "args": {
      "External id": 392,      "device": 2, "stream": 0,      "correlation": 203, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 203, "pid": 2, "tid": 0, "ts": 7253740886001.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902507.939, "dur": 3.279,
    "args": {
      "External id": 976,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 498,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 498, "pid": 49625, "tid": 49625, "ts": 7253740902507.939,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886081.246, "dur": 65.509,
    "args": {
      "External id": 393,      "device": 2, "stream": 0,      "correlation": 204, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 204, "pid": 2, "tid": 0, "ts": 7253740886081.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902516.354, "dur": 3.182,
    "args": {
      "External id": 977,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 499,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 499, "pid": 49625, "tid": 49625, "ts": 7253740902516.354,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740886146.756, "dur": 79.662,
    "args": {
      "External id": 395,      "device": 2, "stream": 0,      "correlation": 205, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 205, "pid": 2, "tid": 0, "ts": 7253740886146.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902525.936, "dur": 3.218,
    "args": {
      "External id": 978,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 500,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 500, "pid": 49625, "tid": 49625, "ts": 7253740902525.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740886226.419, "dur": 87.360,
    "args": {
      "External id": 396,      "device": 2, "stream": 0,      "correlation": 206, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 206, "pid": 2, "tid": 0, "ts": 7253740886226.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740886313.780, "dur": 57.210,
    "args": {
      "External id": 399,      "device": 2, "stream": 0,      "correlation": 207, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2346.012619950709
    }
  },
  {
    "ph": "f", "id": 207, "pid": 2, "tid": 0, "ts": 7253740886313.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886370.991, "dur": 6.053,
    "args": {
      "External id": 406,      "device": 2, "stream": 0,      "correlation": 208, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 208, "pid": 2, "tid": 0, "ts": 7253740886370.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902533.952, "dur": 3.220,
    "args": {
      "External id": 980,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 501,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 501, "pid": 49625, "tid": 49625, "ts": 7253740902533.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886377.045, "dur": 5.412,
    "args": {
      "External id": 407,      "device": 2, "stream": 0,      "correlation": 209, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 209, "pid": 2, "tid": 0, "ts": 7253740886377.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902540.896, "dur": 3.274,
    "args": {
      "External id": 981,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 502,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 502, "pid": 49625, "tid": 49625, "ts": 7253740902540.896,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740886382.458, "dur": 87.640,
    "args": {
      "External id": 408,      "device": 2, "stream": 0,      "correlation": 210, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 210, "pid": 2, "tid": 0, "ts": 7253740886382.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886470.099, "dur": 63.906,
    "args": {
      "External id": 409,      "device": 2, "stream": 0,      "correlation": 211, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 211, "pid": 2, "tid": 0, "ts": 7253740886470.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740886534.006, "dur": 79.501,
    "args": {
      "External id": 410,      "device": 2, "stream": 0,      "correlation": 212, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 212, "pid": 2, "tid": 0, "ts": 7253740886534.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886613.508, "dur": 64.548,
    "args": {
      "External id": 411,      "device": 2, "stream": 0,      "correlation": 213, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 213, "pid": 2, "tid": 0, "ts": 7253740886613.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740886678.057, "dur": 79.220,
    "args": {
      "External id": 413,      "device": 2, "stream": 0,      "correlation": 214, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 214, "pid": 2, "tid": 0, "ts": 7253740886678.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740886757.278, "dur": 87.601,
    "args": {
      "External id": 414,      "device": 2, "stream": 0,      "correlation": 215, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 215, "pid": 2, "tid": 0, "ts": 7253740886757.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740886844.880, "dur": 61.420,
    "args": {
      "External id": 417,      "device": 2, "stream": 0,      "correlation": 216, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2185.209097865551
    }
  },
  {
    "ph": "f", "id": 216, "pid": 2, "tid": 0, "ts": 7253740886844.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886906.301, "dur": 6.534,
    "args": {
      "External id": 424,      "device": 2, "stream": 0,      "correlation": 217, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 217, "pid": 2, "tid": 0, "ts": 7253740886906.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902570.794, "dur": 3.419,
    "args": {
      "External id": 983,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 503,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 503, "pid": 49625, "tid": 49625, "ts": 7253740902570.794,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740886912.836, "dur": 8.939,
    "args": {
      "External id": 425,      "device": 2, "stream": 0,      "correlation": 218, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 218, "pid": 2, "tid": 0, "ts": 7253740886912.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740886921.776, "dur": 87.841,
    "args": {
      "External id": 426,      "device": 2, "stream": 0,      "correlation": 219, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 219, "pid": 2, "tid": 0, "ts": 7253740886921.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887009.618, "dur": 63.024,
    "args": {
      "External id": 427,      "device": 2, "stream": 0,      "correlation": 220, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 220, "pid": 2, "tid": 0, "ts": 7253740887009.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740887072.643, "dur": 79.341,
    "args": {
      "External id": 428,      "device": 2, "stream": 0,      "correlation": 221, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 221, "pid": 2, "tid": 0, "ts": 7253740887072.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902583.432, "dur": 3.768,
    "args": {
      "External id": 986,      "cid": 134, "correlation": 504, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 504, "pid": 49625, "tid": 49625, "ts": 7253740902583.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887151.985, "dur": 64.307,
    "args": {
      "External id": 429,      "device": 2, "stream": 0,      "correlation": 222, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 222, "pid": 2, "tid": 0, "ts": 7253740887151.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740887216.293, "dur": 80.263,
    "args": {
      "External id": 431,      "device": 2, "stream": 0,      "correlation": 223, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 223, "pid": 2, "tid": 0, "ts": 7253740887216.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740887296.557, "dur": 88.001,
    "args": {
      "External id": 432,      "device": 2, "stream": 0,      "correlation": 224, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 224, "pid": 2, "tid": 0, "ts": 7253740887296.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740887384.559, "dur": 59.295,
    "args": {
      "External id": 435,      "device": 2, "stream": 0,      "correlation": 225, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2263.5207771181867
    }
  },
  {
    "ph": "f", "id": 225, "pid": 2, "tid": 0, "ts": 7253740887384.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887443.855, "dur": 6.093,
    "args": {
      "External id": 442,      "device": 2, "stream": 0,      "correlation": 226, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 226, "pid": 2, "tid": 0, "ts": 7253740887443.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887449.949, "dur": 5.733,
    "args": {
      "External id": 443,      "device": 2, "stream": 0,      "correlation": 227, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 227, "pid": 2, "tid": 0, "ts": 7253740887449.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740887455.683, "dur": 87.760,
    "args": {
      "External id": 444,      "device": 2, "stream": 0,      "correlation": 228, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 228, "pid": 2, "tid": 0, "ts": 7253740887455.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887543.444, "dur": 60.418,
    "args": {
      "External id": 445,      "device": 2, "stream": 0,      "correlation": 229, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 229, "pid": 2, "tid": 0, "ts": 7253740887543.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740887603.863, "dur": 80.504,
    "args": {
      "External id": 446,      "device": 2, "stream": 0,      "correlation": 230, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 230, "pid": 2, "tid": 0, "ts": 7253740887603.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887684.368, "dur": 65.309,
    "args": {
      "External id": 447,      "device": 2, "stream": 0,      "correlation": 231, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 231, "pid": 2, "tid": 0, "ts": 7253740887684.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740887749.678, "dur": 79.141,
    "args": {
      "External id": 449,      "device": 2, "stream": 0,      "correlation": 232, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 232, "pid": 2, "tid": 0, "ts": 7253740887749.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740887828.820, "dur": 87.840,
    "args": {
      "External id": 450,      "device": 2, "stream": 0,      "correlation": 233, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 233, "pid": 2, "tid": 0, "ts": 7253740887828.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740887916.661, "dur": 62.383,
    "args": {
      "External id": 453,      "device": 2, "stream": 0,      "correlation": 234, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2151.476788920236
    }
  },
  {
    "ph": "f", "id": 234, "pid": 2, "tid": 0, "ts": 7253740887916.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887979.045, "dur": 6.093,
    "args": {
      "External id": 460,      "device": 2, "stream": 0,      "correlation": 235, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 235, "pid": 2, "tid": 0, "ts": 7253740887979.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740887985.139, "dur": 5.852,
    "args": {
      "External id": 461,      "device": 2, "stream": 0,      "correlation": 236, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 236, "pid": 2, "tid": 0, "ts": 7253740887985.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902634.167, "dur": 4.066,
    "args": {
      "External id": 993,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 505,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 505, "pid": 49625, "tid": 49625, "ts": 7253740902634.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740887990.992, "dur": 88.081,
    "args": {
      "External id": 462,      "device": 2, "stream": 0,      "correlation": 237, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 237, "pid": 2, "tid": 0, "ts": 7253740887990.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888079.074, "dur": 61.501,
    "args": {
      "External id": 463,      "device": 2, "stream": 0,      "correlation": 238, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 238, "pid": 2, "tid": 0, "ts": 7253740888079.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740888140.576, "dur": 79.702,
    "args": {
      "External id": 464,      "device": 2, "stream": 0,      "correlation": 239, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 239, "pid": 2, "tid": 0, "ts": 7253740888140.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902643.839, "dur": 4.012,
    "args": {
      "External id": 994,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 506,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 506, "pid": 49625, "tid": 49625, "ts": 7253740902643.839,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888220.279, "dur": 67.474,
    "args": {
      "External id": 465,      "device": 2, "stream": 0,      "correlation": 240, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 240, "pid": 2, "tid": 0, "ts": 7253740888220.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740888287.754, "dur": 79.822,
    "args": {
      "External id": 467,      "device": 2, "stream": 0,      "correlation": 241, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 241, "pid": 2, "tid": 0, "ts": 7253740888287.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902652.600, "dur": 3.303,
    "args": {
      "External id": 995,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 507,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 507, "pid": 49625, "tid": 49625, "ts": 7253740902652.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902659.730, "dur": 2.968,
    "args": {
      "External id": 996,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 508,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 508, "pid": 49625, "tid": 49625, "ts": 7253740902659.730,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902668.255, "dur": 4.249,
    "args": {
      "External id": 997,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 509,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 509, "pid": 49625, "tid": 49625, "ts": 7253740902668.255,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902681.545, "dur": 4.388,
    "args": {
      "External id": 998,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 510,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 510, "pid": 49625, "tid": 49625, "ts": 7253740902681.545,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902691.386, "dur": 3.944,
    "args": {
      "External id": 1000,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 511,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 511, "pid": 49625, "tid": 49625, "ts": 7253740902691.386,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740888367.577, "dur": 88.602,
    "args": {
      "External id": 468,      "device": 2, "stream": 0,      "correlation": 242, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 242, "pid": 2, "tid": 0, "ts": 7253740888367.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740888456.180, "dur": 61.300,
    "args": {
      "External id": 471,      "device": 2, "stream": 0,      "correlation": 243, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2189.4867620430337
    }
  },
  {
    "ph": "f", "id": 243, "pid": 2, "tid": 0, "ts": 7253740888456.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888517.481, "dur": 5.973,
    "args": {
      "External id": 478,      "device": 2, "stream": 0,      "correlation": 244, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 244, "pid": 2, "tid": 0, "ts": 7253740888517.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888523.455, "dur": 6.333,
    "args": {
      "External id": 479,      "device": 2, "stream": 0,      "correlation": 245, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 245, "pid": 2, "tid": 0, "ts": 7253740888523.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740888529.789, "dur": 87.921,
    "args": {
      "External id": 480,      "device": 2, "stream": 0,      "correlation": 246, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 246, "pid": 2, "tid": 0, "ts": 7253740888529.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888617.711, "dur": 62.343,
    "args": {
      "External id": 481,      "device": 2, "stream": 0,      "correlation": 247, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 247, "pid": 2, "tid": 0, "ts": 7253740888617.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902706.515, "dur": 3.952,
    "args": {
      "External id": 1001,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 512,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 512, "pid": 49625, "tid": 49625, "ts": 7253740902706.515,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740888680.055, "dur": 79.942,
    "args": {
      "External id": 482,      "device": 2, "stream": 0,      "correlation": 248, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 248, "pid": 2, "tid": 0, "ts": 7253740888680.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740888759.998, "dur": 67.675,
    "args": {
      "External id": 483,      "device": 2, "stream": 0,      "correlation": 249, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 249, "pid": 2, "tid": 0, "ts": 7253740888759.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740888827.674, "dur": 78.499,
    "args": {
      "External id": 485,      "device": 2, "stream": 0,      "correlation": 250, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 250, "pid": 2, "tid": 0, "ts": 7253740888827.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740888906.174, "dur": 86.758,
    "args": {
      "External id": 486,      "device": 2, "stream": 0,      "correlation": 251, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 251, "pid": 2, "tid": 0, "ts": 7253740888906.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740888992.933, "dur": 61.180,
    "args": {
      "External id": 489,      "device": 2, "stream": 0,      "correlation": 252, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2193.7812065837434
    }
  },
  {
    "ph": "f", "id": 252, "pid": 2, "tid": 0, "ts": 7253740888992.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889054.114, "dur": 5.852,
    "args": {
      "External id": 496,      "device": 2, "stream": 0,      "correlation": 253, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 253, "pid": 2, "tid": 0, "ts": 7253740889054.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889059.967, "dur": 5.692,
    "args": {
      "External id": 497,      "device": 2, "stream": 0,      "correlation": 254, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 254, "pid": 2, "tid": 0, "ts": 7253740889059.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740889065.660, "dur": 87.159,
    "args": {
      "External id": 498,      "device": 2, "stream": 0,      "correlation": 255, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 255, "pid": 2, "tid": 0, "ts": 7253740889065.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889152.820, "dur": 64.347,
    "args": {
      "External id": 499,      "device": 2, "stream": 0,      "correlation": 256, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 256, "pid": 2, "tid": 0, "ts": 7253740889152.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740889217.168, "dur": 79.742,
    "args": {
      "External id": 500,      "device": 2, "stream": 0,      "correlation": 257, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 257, "pid": 2, "tid": 0, "ts": 7253740889217.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889296.911, "dur": 62.824,
    "args": {
      "External id": 501,      "device": 2, "stream": 0,      "correlation": 258, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 258, "pid": 2, "tid": 0, "ts": 7253740889296.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902744.122, "dur": 4.059,
    "args": {
      "External id": 1003,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 513,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 513, "pid": 49625, "tid": 49625, "ts": 7253740902744.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740889359.736, "dur": 78.419,
    "args": {
      "External id": 503,      "device": 2, "stream": 0,      "correlation": 259, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 259, "pid": 2, "tid": 0, "ts": 7253740889359.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740889438.156, "dur": 86.397,
    "args": {
      "External id": 504,      "device": 2, "stream": 0,      "correlation": 260, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 260, "pid": 2, "tid": 0, "ts": 7253740889438.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740889524.554, "dur": 58.293,
    "args": {
      "External id": 507,      "device": 2, "stream": 0,      "correlation": 261, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2302.427831337702
    }
  },
  {
    "ph": "f", "id": 261, "pid": 2, "tid": 0, "ts": 7253740889524.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889582.848, "dur": 6.053,
    "args": {
      "External id": 514,      "device": 2, "stream": 0,      "correlation": 262, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 262, "pid": 2, "tid": 0, "ts": 7253740889582.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889588.902, "dur": 5.732,
    "args": {
      "External id": 515,      "device": 2, "stream": 0,      "correlation": 263, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 263, "pid": 2, "tid": 0, "ts": 7253740889588.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740889594.635, "dur": 86.959,
    "args": {
      "External id": 516,      "device": 2, "stream": 0,      "correlation": 264, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 264, "pid": 2, "tid": 0, "ts": 7253740889594.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902760.148, "dur": 4.135,
    "args": {
      "External id": 1006,      "cid": 134, "correlation": 514, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 514, "pid": 49625, "tid": 49625, "ts": 7253740902760.148,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889681.595, "dur": 58.293,
    "args": {
      "External id": 517,      "device": 2, "stream": 0,      "correlation": 265, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 265, "pid": 2, "tid": 0, "ts": 7253740889681.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740889739.889, "dur": 79.902,
    "args": {
      "External id": 518,      "device": 2, "stream": 0,      "correlation": 266, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 266, "pid": 2, "tid": 0, "ts": 7253740889739.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740889819.792, "dur": 63.585,
    "args": {
      "External id": 519,      "device": 2, "stream": 0,      "correlation": 267, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 267, "pid": 2, "tid": 0, "ts": 7253740889819.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740889883.378, "dur": 78.500,
    "args": {
      "External id": 521,      "device": 2, "stream": 0,      "correlation": 268, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 268, "pid": 2, "tid": 0, "ts": 7253740889883.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740889961.879, "dur": 88.161,
    "args": {
      "External id": 522,      "device": 2, "stream": 0,      "correlation": 269, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 269, "pid": 2, "tid": 0, "ts": 7253740889961.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740890050.041, "dur": 60.057,
    "args": {
      "External id": 525,      "device": 2, "stream": 0,      "correlation": 270, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2234.801824902594
    }
  },
  {
    "ph": "f", "id": 270, "pid": 2, "tid": 0, "ts": 7253740890050.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890110.099, "dur": 5.812,
    "args": {
      "External id": 532,      "device": 2, "stream": 0,      "correlation": 271, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 271, "pid": 2, "tid": 0, "ts": 7253740890110.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890115.912, "dur": 6.053,
    "args": {
      "External id": 533,      "device": 2, "stream": 0,      "correlation": 272, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 272, "pid": 2, "tid": 0, "ts": 7253740890115.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740890121.966, "dur": 87.400,
    "args": {
      "External id": 534,      "device": 2, "stream": 0,      "correlation": 273, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 273, "pid": 2, "tid": 0, "ts": 7253740890121.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890209.367, "dur": 61.901,
    "args": {
      "External id": 535,      "device": 2, "stream": 0,      "correlation": 274, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 274, "pid": 2, "tid": 0, "ts": 7253740890209.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740890271.269, "dur": 79.542,
    "args": {
      "External id": 536,      "device": 2, "stream": 0,      "correlation": 275, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 275, "pid": 2, "tid": 0, "ts": 7253740890271.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890350.812, "dur": 63.304,
    "args": {
      "External id": 537,      "device": 2, "stream": 0,      "correlation": 276, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 276, "pid": 2, "tid": 0, "ts": 7253740890350.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740890414.117, "dur": 79.422,
    "args": {
      "External id": 539,      "device": 2, "stream": 0,      "correlation": 277, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 277, "pid": 2, "tid": 0, "ts": 7253740890414.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740890493.540, "dur": 87.760,
    "args": {
      "External id": 540,      "device": 2, "stream": 0,      "correlation": 278, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 278, "pid": 2, "tid": 0, "ts": 7253740890493.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740890581.301, "dur": 60.418,
    "args": {
      "External id": 543,      "device": 2, "stream": 0,      "correlation": 279, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2221.449014382893
    }
  },
  {
    "ph": "f", "id": 279, "pid": 2, "tid": 0, "ts": 7253740890581.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890641.720, "dur": 5.973,
    "args": {
      "External id": 550,      "device": 2, "stream": 0,      "correlation": 280, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 280, "pid": 2, "tid": 0, "ts": 7253740890641.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890647.694, "dur": 5.692,
    "args": {
      "External id": 551,      "device": 2, "stream": 0,      "correlation": 281, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 281, "pid": 2, "tid": 0, "ts": 7253740890647.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902815.841, "dur": 3.777,
    "args": {
      "External id": 1013,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 515,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 515, "pid": 49625, "tid": 49625, "ts": 7253740902815.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740890653.387, "dur": 87.400,
    "args": {
      "External id": 552,      "device": 2, "stream": 0,      "correlation": 282, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 282, "pid": 2, "tid": 0, "ts": 7253740890653.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890740.788, "dur": 57.491,
    "args": {
      "External id": 553,      "device": 2, "stream": 0,      "correlation": 283, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 283, "pid": 2, "tid": 0, "ts": 7253740890740.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740890798.280, "dur": 78.539,
    "args": {
      "External id": 554,      "device": 2, "stream": 0,      "correlation": 284, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 284, "pid": 2, "tid": 0, "ts": 7253740890798.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740890876.820, "dur": 66.031,
    "args": {
      "External id": 555,      "device": 2, "stream": 0,      "correlation": 285, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 285, "pid": 2, "tid": 0, "ts": 7253740890876.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740890942.852, "dur": 78.740,
    "args": {
      "External id": 557,      "device": 2, "stream": 0,      "correlation": 286, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 286, "pid": 2, "tid": 0, "ts": 7253740890942.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740891021.593, "dur": 85.355,
    "args": {
      "External id": 558,      "device": 2, "stream": 0,      "correlation": 287, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 287, "pid": 2, "tid": 0, "ts": 7253740891021.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740891106.949, "dur": 62.221,
    "args": {
      "External id": 561,      "device": 2, "stream": 0,      "correlation": 288, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2157.078332422616
    }
  },
  {
    "ph": "f", "id": 288, "pid": 2, "tid": 0, "ts": 7253740891106.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902837.781, "dur": 4.021,
    "args": {
      "External id": 1014,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 516,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 516, "pid": 49625, "tid": 49625, "ts": 7253740902837.781,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891169.171, "dur": 6.093,
    "args": {
      "External id": 568,      "device": 2, "stream": 0,      "correlation": 289, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 289, "pid": 2, "tid": 0, "ts": 7253740891169.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891175.265, "dur": 5.733,
    "args": {
      "External id": 569,      "device": 2, "stream": 0,      "correlation": 290, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 290, "pid": 2, "tid": 0, "ts": 7253740891175.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740891180.999, "dur": 85.475,
    "args": {
      "External id": 570,      "device": 2, "stream": 0,      "correlation": 291, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 291, "pid": 2, "tid": 0, "ts": 7253740891180.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902846.866, "dur": 4.544,
    "args": {
      "External id": 1015,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 517,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 517, "pid": 49625, "tid": 49625, "ts": 7253740902846.866,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891266.475, "dur": 62.903,
    "args": {
      "External id": 571,      "device": 2, "stream": 0,      "correlation": 292, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 292, "pid": 2, "tid": 0, "ts": 7253740891266.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740891329.379, "dur": 79.021,
    "args": {
      "External id": 572,      "device": 2, "stream": 0,      "correlation": 293, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 293, "pid": 2, "tid": 0, "ts": 7253740891329.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902855.689, "dur": 3.902,
    "args": {
      "External id": 1016,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 518,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 518, "pid": 49625, "tid": 49625, "ts": 7253740902855.689,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891408.401, "dur": 63.264,
    "args": {
      "External id": 573,      "device": 2, "stream": 0,      "correlation": 294, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 294, "pid": 2, "tid": 0, "ts": 7253740891408.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740891471.666, "dur": 77.938,
    "args": {
      "External id": 575,      "device": 2, "stream": 0,      "correlation": 295, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 295, "pid": 2, "tid": 0, "ts": 7253740891471.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740891549.605, "dur": 86.277,
    "args": {
      "External id": 576,      "device": 2, "stream": 0,      "correlation": 296, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 296, "pid": 2, "tid": 0, "ts": 7253740891549.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740891635.883, "dur": 60.338,
    "args": {
      "External id": 579,      "device": 2, "stream": 0,      "correlation": 297, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2224.3943055072177
    }
  },
  {
    "ph": "f", "id": 297, "pid": 2, "tid": 0, "ts": 7253740891635.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902865.178, "dur": 4.414,
    "args": {
      "External id": 1017,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 519,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 519, "pid": 49625, "tid": 49625, "ts": 7253740902865.178,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891696.222, "dur": 5.973,
    "args": {
      "External id": 586,      "device": 2, "stream": 0,      "correlation": 298, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 298, "pid": 2, "tid": 0, "ts": 7253740891696.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891702.196, "dur": 5.812,
    "args": {
      "External id": 587,      "device": 2, "stream": 0,      "correlation": 299, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 299, "pid": 2, "tid": 0, "ts": 7253740891702.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740891708.009, "dur": 85.155,
    "args": {
      "External id": 588,      "device": 2, "stream": 0,      "correlation": 300, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 300, "pid": 2, "tid": 0, "ts": 7253740891708.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902875.814, "dur": 4.502,
    "args": {
      "External id": 1018,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 520,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 520, "pid": 49625, "tid": 49625, "ts": 7253740902875.814,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891793.165, "dur": 65.108,
    "args": {
      "External id": 589,      "device": 2, "stream": 0,      "correlation": 301, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 301, "pid": 2, "tid": 0, "ts": 7253740891793.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740891858.274, "dur": 78.419,
    "args": {
      "External id": 590,      "device": 2, "stream": 0,      "correlation": 302, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 302, "pid": 2, "tid": 0, "ts": 7253740891858.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902885.539, "dur": 3.350,
    "args": {
      "External id": 1020,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 521,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 521, "pid": 49625, "tid": 49625, "ts": 7253740902885.539,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740891936.694, "dur": 66.873,
    "args": {
      "External id": 591,      "device": 2, "stream": 0,      "correlation": 303, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 303, "pid": 2, "tid": 0, "ts": 7253740891936.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740892003.568, "dur": 76.094,
    "args": {
      "External id": 593,      "device": 2, "stream": 0,      "correlation": 304, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 304, "pid": 2, "tid": 0, "ts": 7253740892003.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902892.902, "dur": 3.447,
    "args": {
      "External id": 1021,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 522,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 522, "pid": 49625, "tid": 49625, "ts": 7253740902892.902,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740892079.663, "dur": 83.551,
    "args": {
      "External id": 594,      "device": 2, "stream": 0,      "correlation": 305, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 305, "pid": 2, "tid": 0, "ts": 7253740892079.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740892163.215, "dur": 61.179,
    "args": {
      "External id": 597,      "device": 2, "stream": 0,      "correlation": 306, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2193.8170644001307
    }
  },
  {
    "ph": "f", "id": 306, "pid": 2, "tid": 0, "ts": 7253740892163.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892224.395, "dur": 5.933,
    "args": {
      "External id": 604,      "device": 2, "stream": 0,      "correlation": 307, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 307, "pid": 2, "tid": 0, "ts": 7253740892224.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892230.329, "dur": 5.692,
    "args": {
      "External id": 605,      "device": 2, "stream": 0,      "correlation": 308, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 308, "pid": 2, "tid": 0, "ts": 7253740892230.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740892236.022, "dur": 83.992,
    "args": {
      "External id": 606,      "device": 2, "stream": 0,      "correlation": 309, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 309, "pid": 2, "tid": 0, "ts": 7253740892236.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892320.015, "dur": 66.632,
    "args": {
      "External id": 607,      "device": 2, "stream": 0,      "correlation": 310, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 310, "pid": 2, "tid": 0, "ts": 7253740892320.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740892386.648, "dur": 77.697,
    "args": {
      "External id": 608,      "device": 2, "stream": 0,      "correlation": 311, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 311, "pid": 2, "tid": 0, "ts": 7253740892386.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892464.346, "dur": 67.915,
    "args": {
      "External id": 609,      "device": 2, "stream": 0,      "correlation": 312, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 312, "pid": 2, "tid": 0, "ts": 7253740892464.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740892532.262, "dur": 76.254,
    "args": {
      "External id": 611,      "device": 2, "stream": 0,      "correlation": 313, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 313, "pid": 2, "tid": 0, "ts": 7253740892532.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740892608.517, "dur": 83.631,
    "args": {
      "External id": 612,      "device": 2, "stream": 0,      "correlation": 314, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 314, "pid": 2, "tid": 0, "ts": 7253740892608.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740892692.149, "dur": 60.218,
    "args": {
      "External id": 615,      "device": 2, "stream": 0,      "correlation": 315, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2228.826915093243
    }
  },
  {
    "ph": "f", "id": 315, "pid": 2, "tid": 0, "ts": 7253740892692.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892752.368, "dur": 5.932,
    "args": {
      "External id": 622,      "device": 2, "stream": 0,      "correlation": 316, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 316, "pid": 2, "tid": 0, "ts": 7253740892752.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740902927.317, "dur": 3.690,
    "args": {
      "External id": 1023,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 523,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 523, "pid": 49625, "tid": 49625, "ts": 7253740902927.317,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892758.301, "dur": 6.013,
    "args": {
      "External id": 623,      "device": 2, "stream": 0,      "correlation": 317, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 317, "pid": 2, "tid": 0, "ts": 7253740892758.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740892764.315, "dur": 84.633,
    "args": {
      "External id": 624,      "device": 2, "stream": 0,      "correlation": 318, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 318, "pid": 2, "tid": 0, "ts": 7253740892764.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892848.949, "dur": 61.661,
    "args": {
      "External id": 625,      "device": 2, "stream": 0,      "correlation": 319, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 319, "pid": 2, "tid": 0, "ts": 7253740892848.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740902964.514, "dur": 4.693,
    "args": {
      "External id": 1026,      "cid": 134, "correlation": 524, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 524, "pid": 49625, "tid": 49625, "ts": 7253740902964.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740892910.611, "dur": 77.176,
    "args": {
      "External id": 626,      "device": 2, "stream": 0,      "correlation": 320, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 320, "pid": 2, "tid": 0, "ts": 7253740892910.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740892987.788, "dur": 63.946,
    "args": {
      "External id": 627,      "device": 2, "stream": 0,      "correlation": 321, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 321, "pid": 2, "tid": 0, "ts": 7253740892987.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740893051.735, "dur": 76.735,
    "args": {
      "External id": 629,      "device": 2, "stream": 0,      "correlation": 322, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 322, "pid": 2, "tid": 0, "ts": 7253740893051.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740893128.471, "dur": 85.396,
    "args": {
      "External id": 630,      "device": 2, "stream": 0,      "correlation": 323, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 323, "pid": 2, "tid": 0, "ts": 7253740893128.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740893213.868, "dur": 62.542,
    "args": {
      "External id": 633,      "device": 2, "stream": 0,      "correlation": 324, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2146.007195049806
    }
  },
  {
    "ph": "f", "id": 324, "pid": 2, "tid": 0, "ts": 7253740893213.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893276.411, "dur": 5.813,
    "args": {
      "External id": 640,      "device": 2, "stream": 0,      "correlation": 325, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 325, "pid": 2, "tid": 0, "ts": 7253740893276.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893282.225, "dur": 5.531,
    "args": {
      "External id": 641,      "device": 2, "stream": 0,      "correlation": 326, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 326, "pid": 2, "tid": 0, "ts": 7253740893282.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740893287.757, "dur": 85.155,
    "args": {
      "External id": 642,      "device": 2, "stream": 0,      "correlation": 327, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 327, "pid": 2, "tid": 0, "ts": 7253740893287.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893372.913, "dur": 63.505,
    "args": {
      "External id": 643,      "device": 2, "stream": 0,      "correlation": 328, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 328, "pid": 2, "tid": 0, "ts": 7253740893372.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740893436.419, "dur": 78.218,
    "args": {
      "External id": 644,      "device": 2, "stream": 0,      "correlation": 329, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 329, "pid": 2, "tid": 0, "ts": 7253740893436.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893514.638, "dur": 69.559,
    "args": {
      "External id": 645,      "device": 2, "stream": 0,      "correlation": 330, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 330, "pid": 2, "tid": 0, "ts": 7253740893514.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740893584.198, "dur": 77.096,
    "args": {
      "External id": 647,      "device": 2, "stream": 0,      "correlation": 331, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 331, "pid": 2, "tid": 0, "ts": 7253740893584.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740893661.295, "dur": 85.476,
    "args": {
      "External id": 648,      "device": 2, "stream": 0,      "correlation": 332, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 332, "pid": 2, "tid": 0, "ts": 7253740893661.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740893746.772, "dur": 61.821,
    "args": {
      "External id": 651,      "device": 2, "stream": 0,      "correlation": 333, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2171.0350360713014
    }
  },
  {
    "ph": "f", "id": 333, "pid": 2, "tid": 0, "ts": 7253740893746.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893808.594, "dur": 5.812,
    "args": {
      "External id": 658,      "device": 2, "stream": 0,      "correlation": 334, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 334, "pid": 2, "tid": 0, "ts": 7253740893808.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893814.407, "dur": 5.973,
    "args": {
      "External id": 659,      "device": 2, "stream": 0,      "correlation": 335, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 335, "pid": 2, "tid": 0, "ts": 7253740893814.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740893820.381, "dur": 84.152,
    "args": {
      "External id": 660,      "device": 2, "stream": 0,      "correlation": 336, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 336, "pid": 2, "tid": 0, "ts": 7253740893820.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903023.789, "dur": 4.369,
    "args": {
      "External id": 1033,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 525,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 525, "pid": 49625, "tid": 49625, "ts": 7253740903023.789,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740893904.534, "dur": 61.861,
    "args": {
      "External id": 661,      "device": 2, "stream": 0,      "correlation": 337, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 337, "pid": 2, "tid": 0, "ts": 7253740893904.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740893966.396, "dur": 77.136,
    "args": {
      "External id": 662,      "device": 2, "stream": 0,      "correlation": 338, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 338, "pid": 2, "tid": 0, "ts": 7253740893966.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894043.533, "dur": 64.427,
    "args": {
      "External id": 663,      "device": 2, "stream": 0,      "correlation": 339, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 339, "pid": 2, "tid": 0, "ts": 7253740894043.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903034.883, "dur": 4.384,
    "args": {
      "External id": 1034,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 526,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 526, "pid": 49625, "tid": 49625, "ts": 7253740903034.883,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740894107.961, "dur": 76.535,
    "args": {
      "External id": 665,      "device": 2, "stream": 0,      "correlation": 340, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 340, "pid": 2, "tid": 0, "ts": 7253740894107.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740894184.497, "dur": 84.393,
    "args": {
      "External id": 666,      "device": 2, "stream": 0,      "correlation": 341, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 341, "pid": 2, "tid": 0, "ts": 7253740894184.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740894268.891, "dur": 59.976,
    "args": {
      "External id": 669,      "device": 2, "stream": 0,      "correlation": 342, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2237.819964319656
    }
  },
  {
    "ph": "f", "id": 342, "pid": 2, "tid": 0, "ts": 7253740894268.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894328.868, "dur": 5.893,
    "args": {
      "External id": 676,      "device": 2, "stream": 0,      "correlation": 343, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 343, "pid": 2, "tid": 0, "ts": 7253740894328.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903044.398, "dur": 4.179,
    "args": {
      "External id": 1035,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 527,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 527, "pid": 49625, "tid": 49625, "ts": 7253740903044.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894334.762, "dur": 5.692,
    "args": {
      "External id": 677,      "device": 2, "stream": 0,      "correlation": 344, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 344, "pid": 2, "tid": 0, "ts": 7253740894334.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740894340.455, "dur": 84.072,
    "args": {
      "External id": 678,      "device": 2, "stream": 0,      "correlation": 345, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 345, "pid": 2, "tid": 0, "ts": 7253740894340.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903052.696, "dur": 3.651,
    "args": {
      "External id": 1036,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 528,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 528, "pid": 49625, "tid": 49625, "ts": 7253740903052.696,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894424.528, "dur": 60.979,
    "args": {
      "External id": 679,      "device": 2, "stream": 0,      "correlation": 346, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 346, "pid": 2, "tid": 0, "ts": 7253740894424.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740894485.508, "dur": 78.379,
    "args": {
      "External id": 680,      "device": 2, "stream": 0,      "correlation": 347, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 347, "pid": 2, "tid": 0, "ts": 7253740894485.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903061.513, "dur": 4.131,
    "args": {
      "External id": 1037,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 529,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 529, "pid": 49625, "tid": 49625, "ts": 7253740903061.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894563.888, "dur": 64.187,
    "args": {
      "External id": 681,      "device": 2, "stream": 0,      "correlation": 348, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 348, "pid": 2, "tid": 0, "ts": 7253740894563.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740894628.076, "dur": 77.617,
    "args": {
      "External id": 683,      "device": 2, "stream": 0,      "correlation": 349, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 349, "pid": 2, "tid": 0, "ts": 7253740894628.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740894705.694, "dur": 84.433,
    "args": {
      "External id": 684,      "device": 2, "stream": 0,      "correlation": 350, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 350, "pid": 2, "tid": 0, "ts": 7253740894705.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740894790.128, "dur": 58.613,
    "args": {
      "External id": 687,      "device": 2, "stream": 0,      "correlation": 351, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2289.8578496604905
    }
  },
  {
    "ph": "f", "id": 351, "pid": 2, "tid": 0, "ts": 7253740894790.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903072.217, "dur": 3.734,
    "args": {
      "External id": 1038,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 530,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 530, "pid": 49625, "tid": 49625, "ts": 7253740903072.217,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894848.742, "dur": 5.733,
    "args": {
      "External id": 694,      "device": 2, "stream": 0,      "correlation": 352, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 352, "pid": 2, "tid": 0, "ts": 7253740894848.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894854.476, "dur": 5.571,
    "args": {
      "External id": 695,      "device": 2, "stream": 0,      "correlation": 353, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 353, "pid": 2, "tid": 0, "ts": 7253740894854.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740894860.048, "dur": 83.150,
    "args": {
      "External id": 696,      "device": 2, "stream": 0,      "correlation": 354, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 354, "pid": 2, "tid": 0, "ts": 7253740894860.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903080.435, "dur": 4.029,
    "args": {
      "External id": 1040,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 531,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 531, "pid": 49625, "tid": 49625, "ts": 7253740903080.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740894943.199, "dur": 66.753,
    "args": {
      "External id": 697,      "device": 2, "stream": 0,      "correlation": 355, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 355, "pid": 2, "tid": 0, "ts": 7253740894943.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740895009.953, "dur": 77.898,
    "args": {
      "External id": 698,      "device": 2, "stream": 0,      "correlation": 356, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 356, "pid": 2, "tid": 0, "ts": 7253740895009.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903087.946, "dur": 3.123,
    "args": {
      "External id": 1041,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 532,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 532, "pid": 49625, "tid": 49625, "ts": 7253740903087.946,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895087.852, "dur": 63.825,
    "args": {
      "External id": 699,      "device": 2, "stream": 0,      "correlation": 357, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 357, "pid": 2, "tid": 0, "ts": 7253740895087.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740895151.678, "dur": 76.094,
    "args": {
      "External id": 701,      "device": 2, "stream": 0,      "correlation": 358, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 358, "pid": 2, "tid": 0, "ts": 7253740895151.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740895227.773, "dur": 83.030,
    "args": {
      "External id": 702,      "device": 2, "stream": 0,      "correlation": 359, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 359, "pid": 2, "tid": 0, "ts": 7253740895227.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740895310.804, "dur": 60.337,
    "args": {
      "External id": 705,      "device": 2, "stream": 0,      "correlation": 360, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2224.431171069641
    }
  },
  {
    "ph": "f", "id": 360, "pid": 2, "tid": 0, "ts": 7253740895310.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895371.142, "dur": 6.454,
    "args": {
      "External id": 712,      "device": 2, "stream": 0,      "correlation": 361, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 361, "pid": 2, "tid": 0, "ts": 7253740895371.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895377.597, "dur": 5.692,
    "args": {
      "External id": 713,      "device": 2, "stream": 0,      "correlation": 362, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 362, "pid": 2, "tid": 0, "ts": 7253740895377.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740895383.290, "dur": 83.311,
    "args": {
      "External id": 714,      "device": 2, "stream": 0,      "correlation": 363, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 363, "pid": 2, "tid": 0, "ts": 7253740895383.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895466.602, "dur": 63.224,
    "args": {
      "External id": 715,      "device": 2, "stream": 0,      "correlation": 364, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 364, "pid": 2, "tid": 0, "ts": 7253740895466.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740895529.827, "dur": 76.775,
    "args": {
      "External id": 716,      "device": 2, "stream": 0,      "correlation": 365, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 365, "pid": 2, "tid": 0, "ts": 7253740895529.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895606.603, "dur": 65.991,
    "args": {
      "External id": 717,      "device": 2, "stream": 0,      "correlation": 366, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 366, "pid": 2, "tid": 0, "ts": 7253740895606.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903121.065, "dur": 4.408,
    "args": {
      "External id": 1043,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 533,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 533, "pid": 49625, "tid": 49625, "ts": 7253740903121.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740895672.595, "dur": 76.174,
    "args": {
      "External id": 719,      "device": 2, "stream": 0,      "correlation": 367, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 367, "pid": 2, "tid": 0, "ts": 7253740895672.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740895748.770, "dur": 81.706,
    "args": {
      "External id": 720,      "device": 2, "stream": 0,      "correlation": 368, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 368, "pid": 2, "tid": 0, "ts": 7253740895748.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740895830.477, "dur": 58.734,
    "args": {
      "External id": 723,      "device": 2, "stream": 0,      "correlation": 369, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2285.140512471269
    }
  },
  {
    "ph": "f", "id": 369, "pid": 2, "tid": 0, "ts": 7253740895830.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895889.212, "dur": 5.853,
    "args": {
      "External id": 730,      "device": 2, "stream": 0,      "correlation": 370, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 370, "pid": 2, "tid": 0, "ts": 7253740895889.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895895.066, "dur": 6.894,
    "args": {
      "External id": 731,      "device": 2, "stream": 0,      "correlation": 371, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 371, "pid": 2, "tid": 0, "ts": 7253740895895.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903135.433, "dur": 3.683,
    "args": {
      "External id": 1046,      "cid": 134, "correlation": 534, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 534, "pid": 49625, "tid": 49625, "ts": 7253740903135.433,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740895901.961, "dur": 80.585,
    "args": {
      "External id": 732,      "device": 2, "stream": 0,      "correlation": 372, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 372, "pid": 2, "tid": 0, "ts": 7253740895901.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740895982.547, "dur": 60.217,
    "args": {
      "External id": 733,      "device": 2, "stream": 0,      "correlation": 373, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 373, "pid": 2, "tid": 0, "ts": 7253740895982.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740896042.765, "dur": 75.733,
    "args": {
      "External id": 734,      "device": 2, "stream": 0,      "correlation": 374, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 374, "pid": 2, "tid": 0, "ts": 7253740896042.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896118.499, "dur": 64.667,
    "args": {
      "External id": 735,      "device": 2, "stream": 0,      "correlation": 375, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 375, "pid": 2, "tid": 0, "ts": 7253740896118.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740896183.167, "dur": 74.130,
    "args": {
      "External id": 737,      "device": 2, "stream": 0,      "correlation": 376, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 376, "pid": 2, "tid": 0, "ts": 7253740896183.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740896257.298, "dur": 80.584,
    "args": {
      "External id": 738,      "device": 2, "stream": 0,      "correlation": 377, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 377, "pid": 2, "tid": 0, "ts": 7253740896257.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740896337.883, "dur": 60.137,
    "args": {
      "External id": 741,      "device": 2, "stream": 0,      "correlation": 378, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2231.8289268016892
    }
  },
  {
    "ph": "f", "id": 378, "pid": 2, "tid": 0, "ts": 7253740896337.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896398.021, "dur": 5.812,
    "args": {
      "External id": 748,      "device": 2, "stream": 0,      "correlation": 379, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 379, "pid": 2, "tid": 0, "ts": 7253740896398.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896403.834, "dur": 5.772,
    "args": {
      "External id": 749,      "device": 2, "stream": 0,      "correlation": 380, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 380, "pid": 2, "tid": 0, "ts": 7253740896403.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740896409.607, "dur": 80.705,
    "args": {
      "External id": 750,      "device": 2, "stream": 0,      "correlation": 381, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 381, "pid": 2, "tid": 0, "ts": 7253740896409.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896490.313, "dur": 64.106,
    "args": {
      "External id": 751,      "device": 2, "stream": 0,      "correlation": 382, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 382, "pid": 2, "tid": 0, "ts": 7253740896490.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740896554.420, "dur": 75.933,
    "args": {
      "External id": 752,      "device": 2, "stream": 0,      "correlation": 383, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 383, "pid": 2, "tid": 0, "ts": 7253740896554.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896630.354, "dur": 67.555,
    "args": {
      "External id": 753,      "device": 2, "stream": 0,      "correlation": 384, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 384, "pid": 2, "tid": 0, "ts": 7253740896630.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740896697.910, "dur": 74.530,
    "args": {
      "External id": 755,      "device": 2, "stream": 0,      "correlation": 385, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 385, "pid": 2, "tid": 0, "ts": 7253740896697.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740896772.441, "dur": 81.065,
    "args": {
      "External id": 756,      "device": 2, "stream": 0,      "correlation": 386, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 386, "pid": 2, "tid": 0, "ts": 7253740896772.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740896853.507, "dur": 57.651,
    "args": {
      "External id": 759,      "device": 2, "stream": 0,      "correlation": 387, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2328.067161590231
    }
  },
  {
    "ph": "f", "id": 387, "pid": 2, "tid": 0, "ts": 7253740896853.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903188.325, "dur": 3.517,
    "args": {
      "External id": 1053,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 535,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 535, "pid": 49625, "tid": 49625, "ts": 7253740903188.325,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896911.159, "dur": 5.692,
    "args": {
      "External id": 766,      "device": 2, "stream": 0,      "correlation": 388, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 388, "pid": 2, "tid": 0, "ts": 7253740896911.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740896916.852, "dur": 5.773,
    "args": {
      "External id": 767,      "device": 2, "stream": 0,      "correlation": 389, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 389, "pid": 2, "tid": 0, "ts": 7253740896916.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740896922.626, "dur": 79.982,
    "args": {
      "External id": 768,      "device": 2, "stream": 0,      "correlation": 390, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 390, "pid": 2, "tid": 0, "ts": 7253740896922.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897002.609, "dur": 60.097,
    "args": {
      "External id": 769,      "device": 2, "stream": 0,      "correlation": 391, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 391, "pid": 2, "tid": 0, "ts": 7253740897002.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903197.897, "dur": 4.682,
    "args": {
      "External id": 1054,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 536,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 536, "pid": 49625, "tid": 49625, "ts": 7253740903197.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740897062.707, "dur": 75.693,
    "args": {
      "External id": 770,      "device": 2, "stream": 0,      "correlation": 392, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 392, "pid": 2, "tid": 0, "ts": 7253740897062.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897138.401, "dur": 65.670,
    "args": {
      "External id": 771,      "device": 2, "stream": 0,      "correlation": 393, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 393, "pid": 2, "tid": 0, "ts": 7253740897138.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740897204.072, "dur": 73.728,
    "args": {
      "External id": 773,      "device": 2, "stream": 0,      "correlation": 394, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 394, "pid": 2, "tid": 0, "ts": 7253740897204.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903207.725, "dur": 4.452,
    "args": {
      "External id": 1055,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 537,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 537, "pid": 49625, "tid": 49625, "ts": 7253740903207.725,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740897277.801, "dur": 79.863,
    "args": {
      "External id": 774,      "device": 2, "stream": 0,      "correlation": 395, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 395, "pid": 2, "tid": 0, "ts": 7253740897277.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740897357.665, "dur": 60.498,
    "args": {
      "External id": 777,      "device": 2, "stream": 0,      "correlation": 396, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2218.5115125869847
    }
  },
  {
    "ph": "f", "id": 396, "pid": 2, "tid": 0, "ts": 7253740897357.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903216.350, "dur": 3.377,
    "args": {
      "External id": 1056,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 538,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 538, "pid": 49625, "tid": 49625, "ts": 7253740903216.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897418.164, "dur": 5.932,
    "args": {
      "External id": 784,      "device": 2, "stream": 0,      "correlation": 397, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 397, "pid": 2, "tid": 0, "ts": 7253740897418.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903224.593, "dur": 3.608,
    "args": {
      "External id": 1057,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 539,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 539, "pid": 49625, "tid": 49625, "ts": 7253740903224.593,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903234.220, "dur": 3.319,
    "args": {
      "External id": 1058,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 540,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 540, "pid": 49625, "tid": 49625, "ts": 7253740903234.220,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903242.493, "dur": 3.855,
    "args": {
      "External id": 1060,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 541,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 541, "pid": 49625, "tid": 49625, "ts": 7253740903242.493,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903249.925, "dur": 2.996,
    "args": {
      "External id": 1061,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 542,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 542, "pid": 49625, "tid": 49625, "ts": 7253740903249.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897424.097, "dur": 5.693,
    "args": {
      "External id": 785,      "device": 2, "stream": 0,      "correlation": 398, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 398, "pid": 2, "tid": 0, "ts": 7253740897424.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740897429.791, "dur": 80.223,
    "args": {
      "External id": 786,      "device": 2, "stream": 0,      "correlation": 399, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 399, "pid": 2, "tid": 0, "ts": 7253740897429.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897510.015, "dur": 62.863,
    "args": {
      "External id": 787,      "device": 2, "stream": 0,      "correlation": 400, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 400, "pid": 2, "tid": 0, "ts": 7253740897510.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740897572.879, "dur": 76.214,
    "args": {
      "External id": 788,      "device": 2, "stream": 0,      "correlation": 401, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 401, "pid": 2, "tid": 0, "ts": 7253740897572.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897649.094, "dur": 63.946,
    "args": {
      "External id": 789,      "device": 2, "stream": 0,      "correlation": 402, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 402, "pid": 2, "tid": 0, "ts": 7253740897649.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740897713.041, "dur": 74.169,
    "args": {
      "External id": 791,      "device": 2, "stream": 0,      "correlation": 403, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 403, "pid": 2, "tid": 0, "ts": 7253740897713.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740897787.211, "dur": 79.743,
    "args": {
      "External id": 792,      "device": 2, "stream": 0,      "correlation": 404, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 404, "pid": 2, "tid": 0, "ts": 7253740897787.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740897866.955, "dur": 58.934,
    "args": {
      "External id": 795,      "device": 2, "stream": 0,      "correlation": 405, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2277.385730041571
    }
  },
  {
    "ph": "f", "id": 405, "pid": 2, "tid": 0, "ts": 7253740897866.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897925.890, "dur": 6.093,
    "args": {
      "External id": 802,      "device": 2, "stream": 0,      "correlation": 406, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 406, "pid": 2, "tid": 0, "ts": 7253740897925.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740897931.984, "dur": 5.933,
    "args": {
      "External id": 803,      "device": 2, "stream": 0,      "correlation": 407, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 407, "pid": 2, "tid": 0, "ts": 7253740897931.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740897937.918, "dur": 80.103,
    "args": {
      "External id": 804,      "device": 2, "stream": 0,      "correlation": 408, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 408, "pid": 2, "tid": 0, "ts": 7253740897937.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903283.995, "dur": 4.660,
    "args": {
      "External id": 1063,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 543,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 543, "pid": 49625, "tid": 49625, "ts": 7253740903283.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898018.022, "dur": 60.457,
    "args": {
      "External id": 805,      "device": 2, "stream": 0,      "correlation": 409, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 409, "pid": 2, "tid": 0, "ts": 7253740898018.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740898078.480, "dur": 74.370,
    "args": {
      "External id": 806,      "device": 2, "stream": 0,      "correlation": 410, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 410, "pid": 2, "tid": 0, "ts": 7253740898078.480,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898152.851, "dur": 65.189,
    "args": {
      "External id": 807,      "device": 2, "stream": 0,      "correlation": 411, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 411, "pid": 2, "tid": 0, "ts": 7253740898152.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903297.780, "dur": 4.173,
    "args": {
      "External id": 1066,      "cid": 134, "correlation": 544, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 544, "pid": 49625, "tid": 49625, "ts": 7253740903297.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740898218.041, "dur": 74.530,
    "args": {
      "External id": 809,      "device": 2, "stream": 0,      "correlation": 412, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 412, "pid": 2, "tid": 0, "ts": 7253740898218.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740898292.572, "dur": 79.943,
    "args": {
      "External id": 810,      "device": 2, "stream": 0,      "correlation": 413, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 413, "pid": 2, "tid": 0, "ts": 7253740898292.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740898372.516, "dur": 59.135,
    "args": {
      "External id": 813,      "device": 2, "stream": 0,      "correlation": 414, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2269.6450216450216
    }
  },
  {
    "ph": "f", "id": 414, "pid": 2, "tid": 0, "ts": 7253740898372.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898431.652, "dur": 5.972,
    "args": {
      "External id": 820,      "device": 2, "stream": 0,      "correlation": 415, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 415, "pid": 2, "tid": 0, "ts": 7253740898431.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898437.625, "dur": 5.933,
    "args": {
      "External id": 821,      "device": 2, "stream": 0,      "correlation": 416, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 416, "pid": 2, "tid": 0, "ts": 7253740898437.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740898443.559, "dur": 80.223,
    "args": {
      "External id": 822,      "device": 2, "stream": 0,      "correlation": 417, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 417, "pid": 2, "tid": 0, "ts": 7253740898443.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898523.783, "dur": 63.585,
    "args": {
      "External id": 823,      "device": 2, "stream": 0,      "correlation": 418, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 418, "pid": 2, "tid": 0, "ts": 7253740898523.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740898587.369, "dur": 75.412,
    "args": {
      "External id": 824,      "device": 2, "stream": 0,      "correlation": 419, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 419, "pid": 2, "tid": 0, "ts": 7253740898587.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898662.782, "dur": 62.383,
    "args": {
      "External id": 825,      "device": 2, "stream": 0,      "correlation": 420, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 420, "pid": 2, "tid": 0, "ts": 7253740898662.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740898725.166, "dur": 73.327,
    "args": {
      "External id": 827,      "device": 2, "stream": 0,      "correlation": 421, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 421, "pid": 2, "tid": 0, "ts": 7253740898725.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740898798.494, "dur": 79.782,
    "args": {
      "External id": 828,      "device": 2, "stream": 0,      "correlation": 422, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 422, "pid": 2, "tid": 0, "ts": 7253740898798.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740898878.277, "dur": 58.333,
    "args": {
      "External id": 831,      "device": 2, "stream": 0,      "correlation": 423, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2300.8490417252374
    }
  },
  {
    "ph": "f", "id": 423, "pid": 2, "tid": 0, "ts": 7253740898878.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898936.611, "dur": 5.813,
    "args": {
      "External id": 838,      "device": 2, "stream": 0,      "correlation": 424, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 424, "pid": 2, "tid": 0, "ts": 7253740898936.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740898942.425, "dur": 5.652,
    "args": {
      "External id": 839,      "device": 2, "stream": 0,      "correlation": 425, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 425, "pid": 2, "tid": 0, "ts": 7253740898942.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740898948.078, "dur": 78.018,
    "args": {
      "External id": 840,      "device": 2, "stream": 0,      "correlation": 426, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 426, "pid": 2, "tid": 0, "ts": 7253740898948.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899026.097, "dur": 59.495,
    "args": {
      "External id": 841,      "device": 2, "stream": 0,      "correlation": 427, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 427, "pid": 2, "tid": 0, "ts": 7253740899026.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740899085.593, "dur": 73.368,
    "args": {
      "External id": 842,      "device": 2, "stream": 0,      "correlation": 428, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 428, "pid": 2, "tid": 0, "ts": 7253740899085.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903352.379, "dur": 4.481,
    "args": {
      "External id": 1073,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 545,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 545, "pid": 49625, "tid": 49625, "ts": 7253740903352.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899158.962, "dur": 64.347,
    "args": {
      "External id": 843,      "device": 2, "stream": 0,      "correlation": 429, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 429, "pid": 2, "tid": 0, "ts": 7253740899158.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740899223.310, "dur": 72.365,
    "args": {
      "External id": 845,      "device": 2, "stream": 0,      "correlation": 430, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 430, "pid": 2, "tid": 0, "ts": 7253740899223.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740899295.676, "dur": 78.579,
    "args": {
      "External id": 846,      "device": 2, "stream": 0,      "correlation": 431, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 431, "pid": 2, "tid": 0, "ts": 7253740899295.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740899374.256, "dur": 60.939,
    "args": {
      "External id": 849,      "device": 2, "stream": 0,      "correlation": 432, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2202.4569740728584
    }
  },
  {
    "ph": "f", "id": 432, "pid": 2, "tid": 0, "ts": 7253740899374.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903362.868, "dur": 3.938,
    "args": {
      "External id": 1074,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 546,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 546, "pid": 49625, "tid": 49625, "ts": 7253740903362.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899435.196, "dur": 5.652,
    "args": {
      "External id": 856,      "device": 2, "stream": 0,      "correlation": 433, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 433, "pid": 2, "tid": 0, "ts": 7253740899435.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899440.849, "dur": 5.612,
    "args": {
      "External id": 857,      "device": 2, "stream": 0,      "correlation": 434, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 434, "pid": 2, "tid": 0, "ts": 7253740899440.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740899446.462, "dur": 77.898,
    "args": {
      "External id": 858,      "device": 2, "stream": 0,      "correlation": 435, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 435, "pid": 2, "tid": 0, "ts": 7253740899446.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903372.595, "dur": 4.275,
    "args": {
      "External id": 1075,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 547,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 547, "pid": 49625, "tid": 49625, "ts": 7253740903372.595,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899524.361, "dur": 62.382,
    "args": {
      "External id": 859,      "device": 2, "stream": 0,      "correlation": 436, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 436, "pid": 2, "tid": 0, "ts": 7253740899524.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903381.321, "dur": 3.663,
    "args": {
      "External id": 1076,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 548,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 548, "pid": 49625, "tid": 49625, "ts": 7253740903381.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740899586.744, "dur": 74.210,
    "args": {
      "External id": 860,      "device": 2, "stream": 0,      "correlation": 437, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 437, "pid": 2, "tid": 0, "ts": 7253740899586.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899660.955, "dur": 65.028,
    "args": {
      "External id": 861,      "device": 2, "stream": 0,      "correlation": 438, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 438, "pid": 2, "tid": 0, "ts": 7253740899660.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740899725.984, "dur": 72.766,
    "args": {
      "External id": 863,      "device": 2, "stream": 0,      "correlation": 439, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 439, "pid": 2, "tid": 0, "ts": 7253740899725.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903390.207, "dur": 3.539,
    "args": {
      "External id": 1077,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 549,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 549, "pid": 49625, "tid": 49625, "ts": 7253740903390.207,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740899798.751, "dur": 77.257,
    "args": {
      "External id": 864,      "device": 2, "stream": 0,      "correlation": 440, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 440, "pid": 2, "tid": 0, "ts": 7253740899798.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740899876.009, "dur": 59.375,
    "args": {
      "External id": 867,      "device": 2, "stream": 0,      "correlation": 441, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2260.4710320668282
    }
  },
  {
    "ph": "f", "id": 441, "pid": 2, "tid": 0, "ts": 7253740899876.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899935.385, "dur": 5.692,
    "args": {
      "External id": 874,      "device": 2, "stream": 0,      "correlation": 442, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 442, "pid": 2, "tid": 0, "ts": 7253740899935.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740899941.078, "dur": 5.492,
    "args": {
      "External id": 875,      "device": 2, "stream": 0,      "correlation": 443, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 443, "pid": 2, "tid": 0, "ts": 7253740899941.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903400.479, "dur": 3.630,
    "args": {
      "External id": 1078,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 550,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 550, "pid": 49625, "tid": 49625, "ts": 7253740903400.479,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740899946.571, "dur": 76.334,
    "args": {
      "External id": 876,      "device": 2, "stream": 0,      "correlation": 444, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 444, "pid": 2, "tid": 0, "ts": 7253740899946.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903408.535, "dur": 3.042,
    "args": {
      "External id": 1080,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 551,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 551, "pid": 49625, "tid": 49625, "ts": 7253740903408.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900022.906, "dur": 62.102,
    "args": {
      "External id": 877,      "device": 2, "stream": 0,      "correlation": 445, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 445, "pid": 2, "tid": 0, "ts": 7253740900022.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740900085.009, "dur": 71.122,
    "args": {
      "External id": 878,      "device": 2, "stream": 0,      "correlation": 446, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 446, "pid": 2, "tid": 0, "ts": 7253740900085.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900156.132, "dur": 70.842,
    "args": {
      "External id": 879,      "device": 2, "stream": 0,      "correlation": 447, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 447, "pid": 2, "tid": 0, "ts": 7253740900156.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903415.564, "dur": 3.095,
    "args": {
      "External id": 1081,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 552,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 552, "pid": 49625, "tid": 49625, "ts": 7253740903415.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740900226.975, "dur": 71.243,
    "args": {
      "External id": 881,      "device": 2, "stream": 0,      "correlation": 448, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 448, "pid": 2, "tid": 0, "ts": 7253740900226.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740900298.219, "dur": 76.254,
    "args": {
      "External id": 882,      "device": 2, "stream": 0,      "correlation": 449, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 449, "pid": 2, "tid": 0, "ts": 7253740900298.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740900374.474, "dur": 60.097,
    "args": {
      "External id": 885,      "device": 2, "stream": 0,      "correlation": 450, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 2233.3143865020465
    }
  },
  {
    "ph": "f", "id": 450, "pid": 2, "tid": 0, "ts": 7253740900374.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900434.572, "dur": 4.930,
    "args": {
      "External id": 892,      "device": 2, "stream": 0,      "correlation": 451, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 451, "pid": 2, "tid": 0, "ts": 7253740900434.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900439.503, "dur": 4.890,
    "args": {
      "External id": 893,      "device": 2, "stream": 0,      "correlation": 452, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 452, "pid": 2, "tid": 0, "ts": 7253740900439.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740900444.394, "dur": 76.415,
    "args": {
      "External id": 894,      "device": 2, "stream": 0,      "correlation": 453, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 453, "pid": 2, "tid": 0, "ts": 7253740900444.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900520.810, "dur": 66.030,
    "args": {
      "External id": 895,      "device": 2, "stream": 0,      "correlation": 454, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 454, "pid": 2, "tid": 0, "ts": 7253740900520.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740900586.841, "dur": 72.726,
    "args": {
      "External id": 896,      "device": 2, "stream": 0,      "correlation": 455, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 455, "pid": 2, "tid": 0, "ts": 7253740900586.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740900659.568, "dur": 67.675,
    "args": {
      "External id": 897,      "device": 2, "stream": 0,      "correlation": 456, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 456, "pid": 2, "tid": 0, "ts": 7253740900659.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740900727.244, "dur": 71.643,
    "args": {
      "External id": 899,      "device": 2, "stream": 0,      "correlation": 457, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 457, "pid": 2, "tid": 0, "ts": 7253740900727.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903448.132, "dur": 4.244,
    "args": {
      "External id": 1083,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 553,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 553, "pid": 49625, "tid": 49625, "ts": 7253740903448.132,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740900798.888, "dur": 75.092,
    "args": {
      "External id": 900,      "device": 2, "stream": 0,      "correlation": 458, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 458, "pid": 2, "tid": 0, "ts": 7253740900798.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740901690.657, "dur": 47.668,
    "args": {
      "External id": 903,      "device": 2, "stream": 0,      "correlation": 463, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 463, "pid": 2, "tid": 0, "ts": 7253740901690.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740901738.326, "dur": 79.782,
    "args": {
      "External id": 906,      "device": 2, "stream": 0,      "correlation": 464, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1682.2847975132547
    }
  },
  {
    "ph": "f", "id": 464, "pid": 2, "tid": 0, "ts": 7253740901738.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740901874.719, "dur": 1.924,
    "args": {
      "External id": 913,      "device": 2, "stream": 0,      "correlation": 465, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 465, "pid": 2, "tid": 0, "ts": 7253740901874.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740901901.541, "dur": 2.685,
    "args": {
      "External id": 914,      "device": 2, "stream": 0,      "correlation": 466, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 466, "pid": 2, "tid": 0, "ts": 7253740901901.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903461.460, "dur": 4.019,
    "args": {
      "External id": 1086,      "cid": 134, "correlation": 554, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 554, "pid": 49625, "tid": 49625, "ts": 7253740903461.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740901916.335, "dur": 76.775,
    "args": {
      "External id": 915,      "device": 2, "stream": 0,      "correlation": 467, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 467, "pid": 2, "tid": 0, "ts": 7253740901916.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740901993.111, "dur": 58.373,
    "args": {
      "External id": 916,      "device": 2, "stream": 0,      "correlation": 468, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 468, "pid": 2, "tid": 0, "ts": 7253740901993.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740902051.485, "dur": 74.651,
    "args": {
      "External id": 917,      "device": 2, "stream": 0,      "correlation": 469, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 469, "pid": 2, "tid": 0, "ts": 7253740902051.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902126.137, "dur": 60.538,
    "args": {
      "External id": 918,      "device": 2, "stream": 0,      "correlation": 470, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 470, "pid": 2, "tid": 0, "ts": 7253740902126.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740902186.676, "dur": 74.971,
    "args": {
      "External id": 920,      "device": 2, "stream": 0,      "correlation": 471, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 471, "pid": 2, "tid": 0, "ts": 7253740902186.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740902261.648, "dur": 72.245,
    "args": {
      "External id": 921,      "device": 2, "stream": 0,      "correlation": 472, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 472, "pid": 2, "tid": 0, "ts": 7253740902261.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902341.231, "dur": 44.782,
    "args": {
      "External id": 923,      "device": 2, "stream": 0,      "correlation": 473, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 473, "pid": 2, "tid": 0, "ts": 7253740902341.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740902386.014, "dur": 77.537,
    "args": {
      "External id": 926,      "device": 2, "stream": 0,      "correlation": 474, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1730.9929067038097
    }
  },
  {
    "ph": "f", "id": 474, "pid": 2, "tid": 0, "ts": 7253740902386.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902471.209, "dur": 5.051,
    "args": {
      "External id": 933,      "device": 2, "stream": 0,      "correlation": 475, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 475, "pid": 2, "tid": 0, "ts": 7253740902471.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902476.261, "dur": 5.291,
    "args": {
      "External id": 934,      "device": 2, "stream": 0,      "correlation": 476, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 476, "pid": 2, "tid": 0, "ts": 7253740902476.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740902481.553, "dur": 75.853,
    "args": {
      "External id": 935,      "device": 2, "stream": 0,      "correlation": 477, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 477, "pid": 2, "tid": 0, "ts": 7253740902481.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902557.407, "dur": 66.833,
    "args": {
      "External id": 936,      "device": 2, "stream": 0,      "correlation": 478, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 478, "pid": 2, "tid": 0, "ts": 7253740902557.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740902624.241, "dur": 76.174,
    "args": {
      "External id": 937,      "device": 2, "stream": 0,      "correlation": 479, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 479, "pid": 2, "tid": 0, "ts": 7253740902624.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903514.933, "dur": 3.765,
    "args": {
      "External id": 1093,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 555,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 555, "pid": 49625, "tid": 49625, "ts": 7253740903514.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903524.098, "dur": 3.054,
    "args": {
      "External id": 1094,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 556,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 556, "pid": 49625, "tid": 49625, "ts": 7253740903524.098,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903531.880, "dur": 3.782,
    "args": {
      "External id": 1095,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 557,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 557, "pid": 49625, "tid": 49625, "ts": 7253740903531.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903539.902, "dur": 3.336,
    "args": {
      "External id": 1096,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 558,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 558, "pid": 49625, "tid": 49625, "ts": 7253740903539.902,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903549.767, "dur": 4.784,
    "args": {
      "External id": 1097,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 559,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 559, "pid": 49625, "tid": 49625, "ts": 7253740903549.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903560.174, "dur": 3.568,
    "args": {
      "External id": 1098,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 560,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 560, "pid": 49625, "tid": 49625, "ts": 7253740903560.174,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903568.164, "dur": 3.324,
    "args": {
      "External id": 1100,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 561,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 561, "pid": 49625, "tid": 49625, "ts": 7253740903568.164,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903575.434, "dur": 3.252,
    "args": {
      "External id": 1101,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 562,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 562, "pid": 49625, "tid": 49625, "ts": 7253740903575.434,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903607.229, "dur": 3.810,
    "args": {
      "External id": 1103,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 563,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 563, "pid": 49625, "tid": 49625, "ts": 7253740903607.229,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903619.302, "dur": 4.318,
    "args": {
      "External id": 1106,      "cid": 134, "correlation": 564, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 564, "pid": 49625, "tid": 49625, "ts": 7253740903619.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903670.589, "dur": 3.544,
    "args": {
      "External id": 1113,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 565,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 565, "pid": 49625, "tid": 49625, "ts": 7253740903670.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903679.897, "dur": 3.488,
    "args": {
      "External id": 1114,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 566,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 566, "pid": 49625, "tid": 49625, "ts": 7253740903679.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903688.085, "dur": 3.647,
    "args": {
      "External id": 1115,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 567,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 567, "pid": 49625, "tid": 49625, "ts": 7253740903688.085,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903695.377, "dur": 3.196,
    "args": {
      "External id": 1116,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 568,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 568, "pid": 49625, "tid": 49625, "ts": 7253740903695.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903703.141, "dur": 3.440,
    "args": {
      "External id": 1117,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 569,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 569, "pid": 49625, "tid": 49625, "ts": 7253740903703.141,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903711.711, "dur": 3.228,
    "args": {
      "External id": 1118,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 570,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 570, "pid": 49625, "tid": 49625, "ts": 7253740903711.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903719.359, "dur": 3.872,
    "args": {
      "External id": 1120,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 571,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 571, "pid": 49625, "tid": 49625, "ts": 7253740903719.359,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903726.716, "dur": 3.510,
    "args": {
      "External id": 1121,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 572,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 572, "pid": 49625, "tid": 49625, "ts": 7253740903726.716,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903757.993, "dur": 4.219,
    "args": {
      "External id": 1123,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 573,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 573, "pid": 49625, "tid": 49625, "ts": 7253740903757.993,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903770.557, "dur": 3.972,
    "args": {
      "External id": 1126,      "cid": 134, "correlation": 574, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 574, "pid": 49625, "tid": 49625, "ts": 7253740903770.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903820.103, "dur": 3.524,
    "args": {
      "External id": 1133,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 575,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 575, "pid": 49625, "tid": 49625, "ts": 7253740903820.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903829.257, "dur": 3.134,
    "args": {
      "External id": 1134,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 576,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 576, "pid": 49625, "tid": 49625, "ts": 7253740903829.257,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903837.169, "dur": 3.598,
    "args": {
      "External id": 1135,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 577,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 577, "pid": 49625, "tid": 49625, "ts": 7253740903837.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903844.514, "dur": 3.231,
    "args": {
      "External id": 1136,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 578,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 578, "pid": 49625, "tid": 49625, "ts": 7253740903844.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903852.188, "dur": 3.528,
    "args": {
      "External id": 1137,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 579,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 579, "pid": 49625, "tid": 49625, "ts": 7253740903852.188,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903861.298, "dur": 3.625,
    "args": {
      "External id": 1138,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 580,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 580, "pid": 49625, "tid": 49625, "ts": 7253740903861.298,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903870.659, "dur": 3.398,
    "args": {
      "External id": 1140,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 581,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 581, "pid": 49625, "tid": 49625, "ts": 7253740903870.659,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903877.994, "dur": 3.440,
    "args": {
      "External id": 1141,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 582,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 582, "pid": 49625, "tid": 49625, "ts": 7253740903877.994,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903908.095, "dur": 3.698,
    "args": {
      "External id": 1143,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 583,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 583, "pid": 49625, "tid": 49625, "ts": 7253740903908.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740903920.316, "dur": 3.945,
    "args": {
      "External id": 1146,      "cid": 134, "correlation": 584, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 584, "pid": 49625, "tid": 49625, "ts": 7253740903920.316,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903968.319, "dur": 19.633,
    "args": {
      "External id": 1153,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 585,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 585, "pid": 49625, "tid": 49625, "ts": 7253740903968.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740903993.957, "dur": 3.158,
    "args": {
      "External id": 1154,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 586,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 586, "pid": 49625, "tid": 49625, "ts": 7253740903993.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904001.796, "dur": 3.515,
    "args": {
      "External id": 1155,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 587,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 587, "pid": 49625, "tid": 49625, "ts": 7253740904001.796,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904009.275, "dur": 3.674,
    "args": {
      "External id": 1156,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 588,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 588, "pid": 49625, "tid": 49625, "ts": 7253740904009.275,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904017.614, "dur": 3.153,
    "args": {
      "External id": 1157,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 589,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 589, "pid": 49625, "tid": 49625, "ts": 7253740904017.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904026.315, "dur": 3.350,
    "args": {
      "External id": 1158,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 590,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 590, "pid": 49625, "tid": 49625, "ts": 7253740904026.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904034.117, "dur": 3.240,
    "args": {
      "External id": 1160,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 591,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 591, "pid": 49625, "tid": 49625, "ts": 7253740904034.117,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904041.344, "dur": 3.231,
    "args": {
      "External id": 1161,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 592,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 592, "pid": 49625, "tid": 49625, "ts": 7253740904041.344,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904071.078, "dur": 5.019,
    "args": {
      "External id": 1163,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 593,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 593, "pid": 49625, "tid": 49625, "ts": 7253740904071.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904085.732, "dur": 3.827,
    "args": {
      "External id": 1166,      "cid": 134, "correlation": 594, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 594, "pid": 49625, "tid": 49625, "ts": 7253740904085.732,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904135.494, "dur": 4.009,
    "args": {
      "External id": 1173,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 595,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 595, "pid": 49625, "tid": 49625, "ts": 7253740904135.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904145.218, "dur": 3.692,
    "args": {
      "External id": 1174,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 596,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 596, "pid": 49625, "tid": 49625, "ts": 7253740904145.218,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904153.574, "dur": 4.099,
    "args": {
      "External id": 1175,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 597,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 597, "pid": 49625, "tid": 49625, "ts": 7253740904153.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904161.679, "dur": 3.850,
    "args": {
      "External id": 1176,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 598,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 598, "pid": 49625, "tid": 49625, "ts": 7253740904161.679,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904170.056, "dur": 3.853,
    "args": {
      "External id": 1177,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 599,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 599, "pid": 49625, "tid": 49625, "ts": 7253740904170.056,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904180.487, "dur": 3.263,
    "args": {
      "External id": 1178,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 600,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 600, "pid": 49625, "tid": 49625, "ts": 7253740904180.487,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904188.133, "dur": 2.901,
    "args": {
      "External id": 1180,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 601,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 601, "pid": 49625, "tid": 49625, "ts": 7253740904188.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904194.607, "dur": 3.445,
    "args": {
      "External id": 1181,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 602,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 602, "pid": 49625, "tid": 49625, "ts": 7253740904194.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904226.915, "dur": 4.009,
    "args": {
      "External id": 1183,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 603,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 603, "pid": 49625, "tid": 49625, "ts": 7253740904226.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904239.606, "dur": 3.812,
    "args": {
      "External id": 1186,      "cid": 134, "correlation": 604, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 604, "pid": 49625, "tid": 49625, "ts": 7253740904239.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904288.031, "dur": 3.373,
    "args": {
      "External id": 1193,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 605,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 605, "pid": 49625, "tid": 49625, "ts": 7253740904288.031,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904296.830, "dur": 3.389,
    "args": {
      "External id": 1194,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 606,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 606, "pid": 49625, "tid": 49625, "ts": 7253740904296.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904304.952, "dur": 3.606,
    "args": {
      "External id": 1195,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 607,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 607, "pid": 49625, "tid": 49625, "ts": 7253740904304.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904312.388, "dur": 3.280,
    "args": {
      "External id": 1196,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 608,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 608, "pid": 49625, "tid": 49625, "ts": 7253740904312.388,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904320.207, "dur": 3.350,
    "args": {
      "External id": 1197,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 609,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 609, "pid": 49625, "tid": 49625, "ts": 7253740904320.207,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904328.881, "dur": 3.046,
    "args": {
      "External id": 1198,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 610,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 610, "pid": 49625, "tid": 49625, "ts": 7253740904328.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904336.161, "dur": 3.718,
    "args": {
      "External id": 1200,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 611,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 611, "pid": 49625, "tid": 49625, "ts": 7253740904336.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904343.351, "dur": 2.889,
    "args": {
      "External id": 1201,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 612,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 612, "pid": 49625, "tid": 49625, "ts": 7253740904343.351,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904374.193, "dur": 3.248,
    "args": {
      "External id": 1203,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 613,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 613, "pid": 49625, "tid": 49625, "ts": 7253740904374.193,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904385.717, "dur": 3.388,
    "args": {
      "External id": 1206,      "cid": 134, "correlation": 614, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 614, "pid": 49625, "tid": 49625, "ts": 7253740904385.717,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904435.371, "dur": 3.432,
    "args": {
      "External id": 1213,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 615,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 615, "pid": 49625, "tid": 49625, "ts": 7253740904435.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904444.179, "dur": 2.990,
    "args": {
      "External id": 1214,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 616,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 616, "pid": 49625, "tid": 49625, "ts": 7253740904444.179,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904451.837, "dur": 3.292,
    "args": {
      "External id": 1215,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 617,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 617, "pid": 49625, "tid": 49625, "ts": 7253740904451.837,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904458.908, "dur": 2.980,
    "args": {
      "External id": 1216,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 618,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 618, "pid": 49625, "tid": 49625, "ts": 7253740904458.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904466.131, "dur": 3.070,
    "args": {
      "External id": 1217,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 619,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 619, "pid": 49625, "tid": 49625, "ts": 7253740904466.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904474.302, "dur": 3.287,
    "args": {
      "External id": 1218,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 620,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 620, "pid": 49625, "tid": 49625, "ts": 7253740904474.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904481.615, "dur": 2.926,
    "args": {
      "External id": 1220,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 621,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 621, "pid": 49625, "tid": 49625, "ts": 7253740904481.615,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904487.978, "dur": 3.018,
    "args": {
      "External id": 1221,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 622,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 622, "pid": 49625, "tid": 49625, "ts": 7253740904487.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904516.466, "dur": 3.393,
    "args": {
      "External id": 1223,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 623,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 623, "pid": 49625, "tid": 49625, "ts": 7253740904516.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904528.140, "dur": 3.571,
    "args": {
      "External id": 1226,      "cid": 134, "correlation": 624, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 624, "pid": 49625, "tid": 49625, "ts": 7253740904528.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904575.222, "dur": 3.993,
    "args": {
      "External id": 1233,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 625,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 625, "pid": 49625, "tid": 49625, "ts": 7253740904575.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904584.847, "dur": 3.429,
    "args": {
      "External id": 1234,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 626,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 626, "pid": 49625, "tid": 49625, "ts": 7253740904584.847,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904592.841, "dur": 3.447,
    "args": {
      "External id": 1235,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 627,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 627, "pid": 49625, "tid": 49625, "ts": 7253740904592.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904599.931, "dur": 3.118,
    "args": {
      "External id": 1236,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 628,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 628, "pid": 49625, "tid": 49625, "ts": 7253740904599.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904607.948, "dur": 3.623,
    "args": {
      "External id": 1237,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 629,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 629, "pid": 49625, "tid": 49625, "ts": 7253740904607.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904616.631, "dur": 3.641,
    "args": {
      "External id": 1238,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 630,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 630, "pid": 49625, "tid": 49625, "ts": 7253740904616.631,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904624.552, "dur": 3.112,
    "args": {
      "External id": 1240,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 631,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 631, "pid": 49625, "tid": 49625, "ts": 7253740904624.552,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904631.013, "dur": 3.884,
    "args": {
      "External id": 1241,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 632,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 632, "pid": 49625, "tid": 49625, "ts": 7253740904631.013,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904661.520, "dur": 3.693,
    "args": {
      "External id": 1243,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 633,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 633, "pid": 49625, "tid": 49625, "ts": 7253740904661.520,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904674.072, "dur": 3.661,
    "args": {
      "External id": 1246,      "cid": 134, "correlation": 634, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 634, "pid": 49625, "tid": 49625, "ts": 7253740904674.072,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904722.905, "dur": 3.681,
    "args": {
      "External id": 1253,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 635,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 635, "pid": 49625, "tid": 49625, "ts": 7253740904722.905,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904731.936, "dur": 3.408,
    "args": {
      "External id": 1254,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 636,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 636, "pid": 49625, "tid": 49625, "ts": 7253740904731.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904740.045, "dur": 4.251,
    "args": {
      "External id": 1255,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 637,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 637, "pid": 49625, "tid": 49625, "ts": 7253740904740.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904748.108, "dur": 3.400,
    "args": {
      "External id": 1256,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 638,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 638, "pid": 49625, "tid": 49625, "ts": 7253740904748.108,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904755.737, "dur": 3.908,
    "args": {
      "External id": 1257,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 639,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 639, "pid": 49625, "tid": 49625, "ts": 7253740904755.737,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904765.019, "dur": 3.505,
    "args": {
      "External id": 1258,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 640,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 640, "pid": 49625, "tid": 49625, "ts": 7253740904765.019,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904773.078, "dur": 3.387,
    "args": {
      "External id": 1260,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 641,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 641, "pid": 49625, "tid": 49625, "ts": 7253740904773.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904779.944, "dur": 3.133,
    "args": {
      "External id": 1261,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 642,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 642, "pid": 49625, "tid": 49625, "ts": 7253740904779.944,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904808.617, "dur": 4.037,
    "args": {
      "External id": 1263,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 643,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 643, "pid": 49625, "tid": 49625, "ts": 7253740904808.617,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904820.721, "dur": 3.387,
    "args": {
      "External id": 1266,      "cid": 134, "correlation": 644, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 644, "pid": 49625, "tid": 49625, "ts": 7253740904820.721,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904884.218, "dur": 3.546,
    "args": {
      "External id": 1273,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 645,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 645, "pid": 49625, "tid": 49625, "ts": 7253740904884.218,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904895.066, "dur": 3.576,
    "args": {
      "External id": 1274,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 646,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 646, "pid": 49625, "tid": 49625, "ts": 7253740904895.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904903.939, "dur": 4.324,
    "args": {
      "External id": 1275,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 647,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 647, "pid": 49625, "tid": 49625, "ts": 7253740904903.939,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904912.211, "dur": 3.844,
    "args": {
      "External id": 1276,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 648,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 648, "pid": 49625, "tid": 49625, "ts": 7253740904912.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904921.029, "dur": 3.353,
    "args": {
      "External id": 1277,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 649,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 649, "pid": 49625, "tid": 49625, "ts": 7253740904921.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904930.173, "dur": 3.260,
    "args": {
      "External id": 1278,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 650,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 650, "pid": 49625, "tid": 49625, "ts": 7253740904930.173,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904938.171, "dur": 3.464,
    "args": {
      "External id": 1280,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 651,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 651, "pid": 49625, "tid": 49625, "ts": 7253740904938.171,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904945.256, "dur": 2.914,
    "args": {
      "External id": 1281,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 652,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 652, "pid": 49625, "tid": 49625, "ts": 7253740904945.256,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740904976.825, "dur": 4.535,
    "args": {
      "External id": 1283,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 653,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 653, "pid": 49625, "tid": 49625, "ts": 7253740904976.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740904990.932, "dur": 3.958,
    "args": {
      "External id": 1286,      "cid": 134, "correlation": 654, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 654, "pid": 49625, "tid": 49625, "ts": 7253740904990.932,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905041.423, "dur": 3.631,
    "args": {
      "External id": 1293,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 655,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 655, "pid": 49625, "tid": 49625, "ts": 7253740905041.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905051.155, "dur": 3.030,
    "args": {
      "External id": 1294,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 656,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 656, "pid": 49625, "tid": 49625, "ts": 7253740905051.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905058.889, "dur": 4.222,
    "args": {
      "External id": 1295,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 657,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 657, "pid": 49625, "tid": 49625, "ts": 7253740905058.889,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905067.084, "dur": 4.357,
    "args": {
      "External id": 1296,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 658,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 658, "pid": 49625, "tid": 49625, "ts": 7253740905067.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905075.888, "dur": 3.342,
    "args": {
      "External id": 1297,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 659,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 659, "pid": 49625, "tid": 49625, "ts": 7253740905075.888,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905084.426, "dur": 3.534,
    "args": {
      "External id": 1298,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 660,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 660, "pid": 49625, "tid": 49625, "ts": 7253740905084.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905092.198, "dur": 3.156,
    "args": {
      "External id": 1300,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 661,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 661, "pid": 49625, "tid": 49625, "ts": 7253740905092.198,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905098.751, "dur": 3.312,
    "args": {
      "External id": 1301,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 662,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 662, "pid": 49625, "tid": 49625, "ts": 7253740905098.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905130.631, "dur": 3.762,
    "args": {
      "External id": 1303,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 663,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 663, "pid": 49625, "tid": 49625, "ts": 7253740905130.631,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905142.779, "dur": 3.603,
    "args": {
      "External id": 1306,      "cid": 134, "correlation": 664, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 664, "pid": 49625, "tid": 49625, "ts": 7253740905142.779,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905191.315, "dur": 3.948,
    "args": {
      "External id": 1313,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 665,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 665, "pid": 49625, "tid": 49625, "ts": 7253740905191.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905200.758, "dur": 3.270,
    "args": {
      "External id": 1314,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 666,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 666, "pid": 49625, "tid": 49625, "ts": 7253740905200.758,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905208.724, "dur": 3.359,
    "args": {
      "External id": 1315,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 667,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 667, "pid": 49625, "tid": 49625, "ts": 7253740905208.724,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905215.575, "dur": 3.181,
    "args": {
      "External id": 1316,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 668,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 668, "pid": 49625, "tid": 49625, "ts": 7253740905215.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905225.497, "dur": 6.374,
    "args": {
      "External id": 1317,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 669,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 669, "pid": 49625, "tid": 49625, "ts": 7253740905225.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905237.903, "dur": 4.048,
    "args": {
      "External id": 1318,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 670,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 670, "pid": 49625, "tid": 49625, "ts": 7253740905237.903,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905246.744, "dur": 3.403,
    "args": {
      "External id": 1320,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 671,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 671, "pid": 49625, "tid": 49625, "ts": 7253740905246.744,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905253.717, "dur": 3.198,
    "args": {
      "External id": 1321,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 672,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 672, "pid": 49625, "tid": 49625, "ts": 7253740905253.717,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905283.241, "dur": 4.789,
    "args": {
      "External id": 1323,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 673,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 673, "pid": 49625, "tid": 49625, "ts": 7253740905283.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905295.848, "dur": 3.593,
    "args": {
      "External id": 1326,      "cid": 134, "correlation": 674, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 674, "pid": 49625, "tid": 49625, "ts": 7253740905295.848,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905340.169, "dur": 3.482,
    "args": {
      "External id": 1333,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 675,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 675, "pid": 49625, "tid": 49625, "ts": 7253740905340.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905348.611, "dur": 3.569,
    "args": {
      "External id": 1334,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 676,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 676, "pid": 49625, "tid": 49625, "ts": 7253740905348.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905356.500, "dur": 3.644,
    "args": {
      "External id": 1335,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 677,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 677, "pid": 49625, "tid": 49625, "ts": 7253740905356.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905363.460, "dur": 3.227,
    "args": {
      "External id": 1336,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 678,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 678, "pid": 49625, "tid": 49625, "ts": 7253740905363.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905370.701, "dur": 4.608,
    "args": {
      "External id": 1337,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 679,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 679, "pid": 49625, "tid": 49625, "ts": 7253740905370.701,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905380.548, "dur": 3.441,
    "args": {
      "External id": 1338,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 680,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 680, "pid": 49625, "tid": 49625, "ts": 7253740905380.548,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905387.803, "dur": 3.278,
    "args": {
      "External id": 1340,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 681,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 681, "pid": 49625, "tid": 49625, "ts": 7253740905387.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905394.367, "dur": 3.308,
    "args": {
      "External id": 1341,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 682,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 682, "pid": 49625, "tid": 49625, "ts": 7253740905394.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905421.947, "dur": 5.180,
    "args": {
      "External id": 1343,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 683,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 683, "pid": 49625, "tid": 49625, "ts": 7253740905421.947,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905434.871, "dur": 4.468,
    "args": {
      "External id": 1346,      "cid": 134, "correlation": 684, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 684, "pid": 49625, "tid": 49625, "ts": 7253740905434.871,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905481.401, "dur": 3.609,
    "args": {
      "External id": 1353,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 685,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 685, "pid": 49625, "tid": 49625, "ts": 7253740905481.401,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905489.901, "dur": 3.496,
    "args": {
      "External id": 1354,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 686,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 686, "pid": 49625, "tid": 49625, "ts": 7253740905489.901,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905497.665, "dur": 3.516,
    "args": {
      "External id": 1355,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 687,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 687, "pid": 49625, "tid": 49625, "ts": 7253740905497.665,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905504.389, "dur": 4.447,
    "args": {
      "External id": 1356,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 688,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 688, "pid": 49625, "tid": 49625, "ts": 7253740905504.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905512.880, "dur": 3.341,
    "args": {
      "External id": 1357,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 689,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 689, "pid": 49625, "tid": 49625, "ts": 7253740905512.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905520.890, "dur": 3.312,
    "args": {
      "External id": 1358,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 690,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 690, "pid": 49625, "tid": 49625, "ts": 7253740905520.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905528.176, "dur": 3.421,
    "args": {
      "External id": 1360,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 691,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 691, "pid": 49625, "tid": 49625, "ts": 7253740905528.176,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905534.662, "dur": 2.786,
    "args": {
      "External id": 1361,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 692,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 692, "pid": 49625, "tid": 49625, "ts": 7253740905534.662,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905561.068, "dur": 4.161,
    "args": {
      "External id": 1363,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 693,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 693, "pid": 49625, "tid": 49625, "ts": 7253740905561.068,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905572.971, "dur": 3.331,
    "args": {
      "External id": 1366,      "cid": 134, "correlation": 694, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 694, "pid": 49625, "tid": 49625, "ts": 7253740905572.971,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905615.990, "dur": 3.203,
    "args": {
      "External id": 1373,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 695,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 695, "pid": 49625, "tid": 49625, "ts": 7253740905615.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905624.121, "dur": 3.033,
    "args": {
      "External id": 1374,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 696,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 696, "pid": 49625, "tid": 49625, "ts": 7253740905624.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905631.321, "dur": 3.030,
    "args": {
      "External id": 1375,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 697,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 697, "pid": 49625, "tid": 49625, "ts": 7253740905631.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905637.601, "dur": 2.992,
    "args": {
      "External id": 1376,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 698,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 698, "pid": 49625, "tid": 49625, "ts": 7253740905637.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905644.507, "dur": 3.197,
    "args": {
      "External id": 1377,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 699,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 699, "pid": 49625, "tid": 49625, "ts": 7253740905644.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905653.530, "dur": 2.927,
    "args": {
      "External id": 1378,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 700,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 700, "pid": 49625, "tid": 49625, "ts": 7253740905653.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905660.293, "dur": 2.830,
    "args": {
      "External id": 1380,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 701,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 701, "pid": 49625, "tid": 49625, "ts": 7253740905660.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905666.272, "dur": 2.918,
    "args": {
      "External id": 1381,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 702,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 702, "pid": 49625, "tid": 49625, "ts": 7253740905666.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905693.214, "dur": 3.755,
    "args": {
      "External id": 1383,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 703,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 703, "pid": 49625, "tid": 49625, "ts": 7253740905693.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905704.458, "dur": 3.435,
    "args": {
      "External id": 1386,      "cid": 134, "correlation": 704, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 704, "pid": 49625, "tid": 49625, "ts": 7253740905704.458,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905749.880, "dur": 3.333,
    "args": {
      "External id": 1393,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 705,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 705, "pid": 49625, "tid": 49625, "ts": 7253740905749.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905758.063, "dur": 2.970,
    "args": {
      "External id": 1394,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 706,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 706, "pid": 49625, "tid": 49625, "ts": 7253740905758.063,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905765.371, "dur": 3.407,
    "args": {
      "External id": 1395,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 707,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 707, "pid": 49625, "tid": 49625, "ts": 7253740905765.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905772.103, "dur": 3.115,
    "args": {
      "External id": 1396,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 708,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 708, "pid": 49625, "tid": 49625, "ts": 7253740905772.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905779.209, "dur": 2.937,
    "args": {
      "External id": 1397,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 709,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 709, "pid": 49625, "tid": 49625, "ts": 7253740905779.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905786.841, "dur": 3.068,
    "args": {
      "External id": 1398,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 710,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 710, "pid": 49625, "tid": 49625, "ts": 7253740905786.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905793.614, "dur": 2.804,
    "args": {
      "External id": 1400,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 711,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 711, "pid": 49625, "tid": 49625, "ts": 7253740905793.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905799.473, "dur": 2.673,
    "args": {
      "External id": 1401,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 712,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 712, "pid": 49625, "tid": 49625, "ts": 7253740905799.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905825.876, "dur": 3.984,
    "args": {
      "External id": 1403,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 713,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 713, "pid": 49625, "tid": 49625, "ts": 7253740905825.876,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905838.434, "dur": 3.422,
    "args": {
      "External id": 1406,      "cid": 134, "correlation": 714, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 714, "pid": 49625, "tid": 49625, "ts": 7253740905838.434,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905882.854, "dur": 3.571,
    "args": {
      "External id": 1413,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 715,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 715, "pid": 49625, "tid": 49625, "ts": 7253740905882.854,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905891.289, "dur": 3.389,
    "args": {
      "External id": 1414,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 716,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 716, "pid": 49625, "tid": 49625, "ts": 7253740905891.289,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905898.854, "dur": 3.167,
    "args": {
      "External id": 1415,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 717,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 717, "pid": 49625, "tid": 49625, "ts": 7253740905898.854,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905905.163, "dur": 3.099,
    "args": {
      "External id": 1416,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 718,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 718, "pid": 49625, "tid": 49625, "ts": 7253740905905.163,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905912.225, "dur": 3.904,
    "args": {
      "External id": 1417,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 719,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 719, "pid": 49625, "tid": 49625, "ts": 7253740905912.225,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905920.895, "dur": 3.207,
    "args": {
      "External id": 1418,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 720,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 720, "pid": 49625, "tid": 49625, "ts": 7253740905920.895,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905927.898, "dur": 3.010,
    "args": {
      "External id": 1420,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 721,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 721, "pid": 49625, "tid": 49625, "ts": 7253740905927.898,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905933.975, "dur": 2.808,
    "args": {
      "External id": 1421,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 722,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 722, "pid": 49625, "tid": 49625, "ts": 7253740905933.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740905961.954, "dur": 3.716,
    "args": {
      "External id": 1423,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 723,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 723, "pid": 49625, "tid": 49625, "ts": 7253740905961.954,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740905973.537, "dur": 3.543,
    "args": {
      "External id": 1426,      "cid": 134, "correlation": 724, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 724, "pid": 49625, "tid": 49625, "ts": 7253740905973.537,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906017.154, "dur": 3.258,
    "args": {
      "External id": 1433,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 725,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 725, "pid": 49625, "tid": 49625, "ts": 7253740906017.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906024.981, "dur": 3.029,
    "args": {
      "External id": 1434,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 726,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 726, "pid": 49625, "tid": 49625, "ts": 7253740906024.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906032.114, "dur": 2.945,
    "args": {
      "External id": 1435,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 727,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 727, "pid": 49625, "tid": 49625, "ts": 7253740906032.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906038.244, "dur": 3.308,
    "args": {
      "External id": 1436,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 728,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 728, "pid": 49625, "tid": 49625, "ts": 7253740906038.244,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906045.481, "dur": 3.192,
    "args": {
      "External id": 1437,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 729,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 729, "pid": 49625, "tid": 49625, "ts": 7253740906045.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906053.338, "dur": 3.295,
    "args": {
      "External id": 1438,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 730,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 730, "pid": 49625, "tid": 49625, "ts": 7253740906053.338,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906060.386, "dur": 3.467,
    "args": {
      "External id": 1440,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 731,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 731, "pid": 49625, "tid": 49625, "ts": 7253740906060.386,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906066.891, "dur": 15.438,
    "args": {
      "External id": 1441,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 732,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 732, "pid": 49625, "tid": 49625, "ts": 7253740906066.891,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906114.027, "dur": 5.044,
    "args": {
      "External id": 1443,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 733,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 733, "pid": 49625, "tid": 49625, "ts": 7253740906114.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906128.395, "dur": 4.435,
    "args": {
      "External id": 1446,      "cid": 134, "correlation": 734, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 734, "pid": 49625, "tid": 49625, "ts": 7253740906128.395,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906181.793, "dur": 3.797,
    "args": {
      "External id": 1453,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 735,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 735, "pid": 49625, "tid": 49625, "ts": 7253740906181.793,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906191.560, "dur": 3.366,
    "args": {
      "External id": 1454,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 736,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 736, "pid": 49625, "tid": 49625, "ts": 7253740906191.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906199.352, "dur": 3.301,
    "args": {
      "External id": 1455,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 737,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 737, "pid": 49625, "tid": 49625, "ts": 7253740906199.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906206.161, "dur": 4.267,
    "args": {
      "External id": 1456,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 738,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 738, "pid": 49625, "tid": 49625, "ts": 7253740906206.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906214.707, "dur": 3.673,
    "args": {
      "External id": 1457,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 739,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 739, "pid": 49625, "tid": 49625, "ts": 7253740906214.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906223.595, "dur": 3.784,
    "args": {
      "External id": 1458,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 740,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 740, "pid": 49625, "tid": 49625, "ts": 7253740906223.595,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906232.099, "dur": 3.876,
    "args": {
      "External id": 1460,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 741,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 741, "pid": 49625, "tid": 49625, "ts": 7253740906232.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906239.156, "dur": 3.218,
    "args": {
      "External id": 1461,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 742,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 742, "pid": 49625, "tid": 49625, "ts": 7253740906239.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906268.700, "dur": 4.423,
    "args": {
      "External id": 1463,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 743,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 743, "pid": 49625, "tid": 49625, "ts": 7253740906268.700,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906280.869, "dur": 4.352,
    "args": {
      "External id": 1466,      "cid": 134, "correlation": 744, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 744, "pid": 49625, "tid": 49625, "ts": 7253740906280.869,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906327.230, "dur": 3.449,
    "args": {
      "External id": 1473,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 745,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 745, "pid": 49625, "tid": 49625, "ts": 7253740906327.230,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906335.673, "dur": 3.010,
    "args": {
      "External id": 1474,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 746,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 746, "pid": 49625, "tid": 49625, "ts": 7253740906335.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906342.730, "dur": 2.779,
    "args": {
      "External id": 1475,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 747,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 747, "pid": 49625, "tid": 49625, "ts": 7253740906342.730,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906348.787, "dur": 2.973,
    "args": {
      "External id": 1476,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 748,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 748, "pid": 49625, "tid": 49625, "ts": 7253740906348.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906355.741, "dur": 2.973,
    "args": {
      "External id": 1477,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 749,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 749, "pid": 49625, "tid": 49625, "ts": 7253740906355.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906363.347, "dur": 2.919,
    "args": {
      "External id": 1478,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 750,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 750, "pid": 49625, "tid": 49625, "ts": 7253740906363.347,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906370.069, "dur": 2.990,
    "args": {
      "External id": 1480,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 751,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 751, "pid": 49625, "tid": 49625, "ts": 7253740906370.069,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906376.022, "dur": 2.929,
    "args": {
      "External id": 1481,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 752,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 752, "pid": 49625, "tid": 49625, "ts": 7253740906376.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906404.606, "dur": 3.371,
    "args": {
      "External id": 1483,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 753,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 753, "pid": 49625, "tid": 49625, "ts": 7253740906404.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906415.763, "dur": 5.217,
    "args": {
      "External id": 1486,      "cid": 134, "correlation": 754, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 754, "pid": 49625, "tid": 49625, "ts": 7253740906415.763,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906463.893, "dur": 3.575,
    "args": {
      "External id": 1493,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 755,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 755, "pid": 49625, "tid": 49625, "ts": 7253740906463.893,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906476.834, "dur": 3.428,
    "args": {
      "External id": 1494,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 756,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 756, "pid": 49625, "tid": 49625, "ts": 7253740906476.834,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906484.425, "dur": 3.188,
    "args": {
      "External id": 1495,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 757,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 757, "pid": 49625, "tid": 49625, "ts": 7253740906484.425,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906490.957, "dur": 3.418,
    "args": {
      "External id": 1496,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 758,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 758, "pid": 49625, "tid": 49625, "ts": 7253740906490.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906498.276, "dur": 3.367,
    "args": {
      "External id": 1497,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 759,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 759, "pid": 49625, "tid": 49625, "ts": 7253740906498.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906506.393, "dur": 7.394,
    "args": {
      "External id": 1498,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 760,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 760, "pid": 49625, "tid": 49625, "ts": 7253740906506.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906517.868, "dur": 3.032,
    "args": {
      "External id": 1500,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 761,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 761, "pid": 49625, "tid": 49625, "ts": 7253740906517.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906524.027, "dur": 3.165,
    "args": {
      "External id": 1501,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 762,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 762, "pid": 49625, "tid": 49625, "ts": 7253740906524.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906553.892, "dur": 4.494,
    "args": {
      "External id": 1503,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 763,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 763, "pid": 49625, "tid": 49625, "ts": 7253740906553.892,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906566.404, "dur": 4.124,
    "args": {
      "External id": 1506,      "cid": 134, "correlation": 764, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 764, "pid": 49625, "tid": 49625, "ts": 7253740906566.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906612.237, "dur": 3.329,
    "args": {
      "External id": 1513,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 765,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 765, "pid": 49625, "tid": 49625, "ts": 7253740906612.237,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906620.527, "dur": 3.050,
    "args": {
      "External id": 1514,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 766,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 766, "pid": 49625, "tid": 49625, "ts": 7253740906620.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906627.722, "dur": 4.856,
    "args": {
      "External id": 1515,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 767,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 767, "pid": 49625, "tid": 49625, "ts": 7253740906627.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906636.131, "dur": 3.629,
    "args": {
      "External id": 1516,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 768,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 768, "pid": 49625, "tid": 49625, "ts": 7253740906636.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906643.767, "dur": 3.084,
    "args": {
      "External id": 1517,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 769,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 769, "pid": 49625, "tid": 49625, "ts": 7253740906643.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906651.437, "dur": 4.434,
    "args": {
      "External id": 1518,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 770,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 770, "pid": 49625, "tid": 49625, "ts": 7253740906651.437,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906659.725, "dur": 3.527,
    "args": {
      "External id": 1520,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 771,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 771, "pid": 49625, "tid": 49625, "ts": 7253740906659.725,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906666.314, "dur": 3.159,
    "args": {
      "External id": 1521,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 772,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 772, "pid": 49625, "tid": 49625, "ts": 7253740906666.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906692.690, "dur": 3.556,
    "args": {
      "External id": 1523,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 773,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 773, "pid": 49625, "tid": 49625, "ts": 7253740906692.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906704.084, "dur": 3.605,
    "args": {
      "External id": 1526,      "cid": 134, "correlation": 774, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 774, "pid": 49625, "tid": 49625, "ts": 7253740906704.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906748.718, "dur": 3.514,
    "args": {
      "External id": 1533,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 775,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 775, "pid": 49625, "tid": 49625, "ts": 7253740906748.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906757.130, "dur": 3.240,
    "args": {
      "External id": 1534,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 776,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 776, "pid": 49625, "tid": 49625, "ts": 7253740906757.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906764.527, "dur": 4.265,
    "args": {
      "External id": 1535,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 777,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 777, "pid": 49625, "tid": 49625, "ts": 7253740906764.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906772.252, "dur": 3.310,
    "args": {
      "External id": 1536,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 778,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 778, "pid": 49625, "tid": 49625, "ts": 7253740906772.252,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906800.581, "dur": 3.257,
    "args": {
      "External id": 1537,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 779,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 779, "pid": 49625, "tid": 49625, "ts": 7253740906800.581,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906808.443, "dur": 4.234,
    "args": {
      "External id": 1538,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 780,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 780, "pid": 49625, "tid": 49625, "ts": 7253740906808.443,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906816.466, "dur": 3.483,
    "args": {
      "External id": 1540,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 781,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 781, "pid": 49625, "tid": 49625, "ts": 7253740906816.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906823.759, "dur": 3.531,
    "args": {
      "External id": 1541,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 782,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 782, "pid": 49625, "tid": 49625, "ts": 7253740906823.759,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906850.970, "dur": 3.605,
    "args": {
      "External id": 1543,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 783,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 783, "pid": 49625, "tid": 49625, "ts": 7253740906850.970,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906862.328, "dur": 3.549,
    "args": {
      "External id": 1546,      "cid": 134, "correlation": 784, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 784, "pid": 49625, "tid": 49625, "ts": 7253740906862.328,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906906.181, "dur": 3.650,
    "args": {
      "External id": 1553,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 785,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 785, "pid": 49625, "tid": 49625, "ts": 7253740906906.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906914.583, "dur": 3.357,
    "args": {
      "External id": 1554,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 786,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 786, "pid": 49625, "tid": 49625, "ts": 7253740906914.583,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906922.284, "dur": 4.402,
    "args": {
      "External id": 1555,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 787,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 787, "pid": 49625, "tid": 49625, "ts": 7253740906922.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906929.999, "dur": 3.463,
    "args": {
      "External id": 1556,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 788,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 788, "pid": 49625, "tid": 49625, "ts": 7253740906929.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906937.578, "dur": 3.334,
    "args": {
      "External id": 1557,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 789,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 789, "pid": 49625, "tid": 49625, "ts": 7253740906937.578,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906945.376, "dur": 4.663,
    "args": {
      "External id": 1558,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 790,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 790, "pid": 49625, "tid": 49625, "ts": 7253740906945.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906953.822, "dur": 3.364,
    "args": {
      "External id": 1560,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 791,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 791, "pid": 49625, "tid": 49625, "ts": 7253740906953.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906960.233, "dur": 3.723,
    "args": {
      "External id": 1561,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 792,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 792, "pid": 49625, "tid": 49625, "ts": 7253740906960.233,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740906987.643, "dur": 3.690,
    "args": {
      "External id": 1563,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 793,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 793, "pid": 49625, "tid": 49625, "ts": 7253740906987.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740906998.859, "dur": 3.946,
    "args": {
      "External id": 1566,      "cid": 134, "correlation": 794, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 794, "pid": 49625, "tid": 49625, "ts": 7253740906998.859,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907046.553, "dur": 3.981,
    "args": {
      "External id": 1573,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 795,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 795, "pid": 49625, "tid": 49625, "ts": 7253740907046.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907055.578, "dur": 3.578,
    "args": {
      "External id": 1574,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 796,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 796, "pid": 49625, "tid": 49625, "ts": 7253740907055.578,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907063.270, "dur": 5.002,
    "args": {
      "External id": 1575,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 797,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 797, "pid": 49625, "tid": 49625, "ts": 7253740907063.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907071.670, "dur": 4.243,
    "args": {
      "External id": 1576,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 798,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 798, "pid": 49625, "tid": 49625, "ts": 7253740907071.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907079.766, "dur": 17.358,
    "args": {
      "External id": 1577,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 799,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 799, "pid": 49625, "tid": 49625, "ts": 7253740907079.766,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907102.199, "dur": 5.126,
    "args": {
      "External id": 1578,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 800,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 800, "pid": 49625, "tid": 49625, "ts": 7253740907102.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907111.708, "dur": 3.137,
    "args": {
      "External id": 1580,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 801,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 801, "pid": 49625, "tid": 49625, "ts": 7253740907111.708,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907117.923, "dur": 4.112,
    "args": {
      "External id": 1581,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 802,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 802, "pid": 49625, "tid": 49625, "ts": 7253740907117.923,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907146.408, "dur": 4.663,
    "args": {
      "External id": 1583,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 803,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 803, "pid": 49625, "tid": 49625, "ts": 7253740907146.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907158.881, "dur": 4.213,
    "args": {
      "External id": 1586,      "cid": 134, "correlation": 804, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 804, "pid": 49625, "tid": 49625, "ts": 7253740907158.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907204.908, "dur": 3.613,
    "args": {
      "External id": 1593,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 805,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 805, "pid": 49625, "tid": 49625, "ts": 7253740907204.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907213.729, "dur": 3.200,
    "args": {
      "External id": 1594,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 806,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 806, "pid": 49625, "tid": 49625, "ts": 7253740907213.729,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907221.177, "dur": 4.135,
    "args": {
      "External id": 1595,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 807,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 807, "pid": 49625, "tid": 49625, "ts": 7253740907221.177,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907228.698, "dur": 3.570,
    "args": {
      "External id": 1596,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 808,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 808, "pid": 49625, "tid": 49625, "ts": 7253740907228.698,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907236.282, "dur": 3.651,
    "args": {
      "External id": 1597,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 809,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 809, "pid": 49625, "tid": 49625, "ts": 7253740907236.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907244.676, "dur": 4.285,
    "args": {
      "External id": 1598,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 810,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 810, "pid": 49625, "tid": 49625, "ts": 7253740907244.676,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907253.009, "dur": 3.591,
    "args": {
      "External id": 1600,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 811,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 811, "pid": 49625, "tid": 49625, "ts": 7253740907253.009,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907259.636, "dur": 2.909,
    "args": {
      "External id": 1601,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 812,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 812, "pid": 49625, "tid": 49625, "ts": 7253740907259.636,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907287.133, "dur": 3.580,
    "args": {
      "External id": 1603,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 813,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 813, "pid": 49625, "tid": 49625, "ts": 7253740907287.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907298.410, "dur": 3.918,
    "args": {
      "External id": 1606,      "cid": 134, "correlation": 814, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 814, "pid": 49625, "tid": 49625, "ts": 7253740907298.410,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907342.049, "dur": 3.206,
    "args": {
      "External id": 1613,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 815,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 815, "pid": 49625, "tid": 49625, "ts": 7253740907342.049,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907349.949, "dur": 3.271,
    "args": {
      "External id": 1614,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 816,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 816, "pid": 49625, "tid": 49625, "ts": 7253740907349.949,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907357.335, "dur": 4.748,
    "args": {
      "External id": 1615,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 817,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 817, "pid": 49625, "tid": 49625, "ts": 7253740907357.335,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907365.484, "dur": 3.245,
    "args": {
      "External id": 1616,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 818,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 818, "pid": 49625, "tid": 49625, "ts": 7253740907365.484,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907373.738, "dur": 3.134,
    "args": {
      "External id": 1617,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 819,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 819, "pid": 49625, "tid": 49625, "ts": 7253740907373.738,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907381.444, "dur": 4.305,
    "args": {
      "External id": 1618,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 820,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 820, "pid": 49625, "tid": 49625, "ts": 7253740907381.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907389.587, "dur": 3.551,
    "args": {
      "External id": 1620,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 821,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 821, "pid": 49625, "tid": 49625, "ts": 7253740907389.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907396.166, "dur": 2.991,
    "args": {
      "External id": 1621,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 822,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 822, "pid": 49625, "tid": 49625, "ts": 7253740907396.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907422.501, "dur": 3.630,
    "args": {
      "External id": 1623,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 823,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 823, "pid": 49625, "tid": 49625, "ts": 7253740907422.501,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907433.601, "dur": 3.387,
    "args": {
      "External id": 1626,      "cid": 134, "correlation": 824, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 824, "pid": 49625, "tid": 49625, "ts": 7253740907433.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907476.830, "dur": 3.380,
    "args": {
      "External id": 1633,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 825,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 825, "pid": 49625, "tid": 49625, "ts": 7253740907476.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907484.962, "dur": 3.081,
    "args": {
      "External id": 1634,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 826,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 826, "pid": 49625, "tid": 49625, "ts": 7253740907484.962,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907492.058, "dur": 4.298,
    "args": {
      "External id": 1635,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 827,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 827, "pid": 49625, "tid": 49625, "ts": 7253740907492.058,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907499.705, "dur": 3.423,
    "args": {
      "External id": 1636,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 828,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 828, "pid": 49625, "tid": 49625, "ts": 7253740907499.705,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907507.222, "dur": 3.277,
    "args": {
      "External id": 1637,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 829,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 829, "pid": 49625, "tid": 49625, "ts": 7253740907507.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907515.086, "dur": 4.003,
    "args": {
      "External id": 1638,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 830,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 830, "pid": 49625, "tid": 49625, "ts": 7253740907515.086,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907522.907, "dur": 3.235,
    "args": {
      "External id": 1640,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 831,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 831, "pid": 49625, "tid": 49625, "ts": 7253740907522.907,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907529.161, "dur": 2.950,
    "args": {
      "External id": 1641,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 832,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 832, "pid": 49625, "tid": 49625, "ts": 7253740907529.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907555.610, "dur": 3.620,
    "args": {
      "External id": 1643,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 833,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 833, "pid": 49625, "tid": 49625, "ts": 7253740907555.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907566.821, "dur": 3.603,
    "args": {
      "External id": 1646,      "cid": 134, "correlation": 834, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 834, "pid": 49625, "tid": 49625, "ts": 7253740907566.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907609.653, "dur": 3.181,
    "args": {
      "External id": 1653,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 835,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 835, "pid": 49625, "tid": 49625, "ts": 7253740907609.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907617.342, "dur": 3.168,
    "args": {
      "External id": 1654,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 836,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 836, "pid": 49625, "tid": 49625, "ts": 7253740907617.342,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907624.974, "dur": 6.766,
    "args": {
      "External id": 1655,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 837,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 837, "pid": 49625, "tid": 49625, "ts": 7253740907624.974,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907635.965, "dur": 3.576,
    "args": {
      "External id": 1656,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 838,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 838, "pid": 49625, "tid": 49625, "ts": 7253740907635.965,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907643.671, "dur": 4.605,
    "args": {
      "External id": 1657,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 839,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 839, "pid": 49625, "tid": 49625, "ts": 7253740907643.671,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902700.416, "dur": 62.302,
    "args": {
      "External id": 938,      "device": 2, "stream": 0,      "correlation": 480, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 480, "pid": 2, "tid": 0, "ts": 7253740902700.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740902762.719, "dur": 74.129,
    "args": {
      "External id": 940,      "device": 2, "stream": 0,      "correlation": 481, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 481, "pid": 2, "tid": 0, "ts": 7253740902762.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907653.209, "dur": 4.748,
    "args": {
      "External id": 1658,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 840,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 840, "pid": 49625, "tid": 49625, "ts": 7253740907653.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740902836.849, "dur": 73.287,
    "args": {
      "External id": 941,      "device": 2, "stream": 0,      "correlation": 482, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 482, "pid": 2, "tid": 0, "ts": 7253740902836.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740902917.474, "dur": 45.023,
    "args": {
      "External id": 943,      "device": 2, "stream": 0,      "correlation": 483, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 483, "pid": 2, "tid": 0, "ts": 7253740902917.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740902962.498, "dur": 80.303,
    "args": {
      "External id": 946,      "device": 2, "stream": 0,      "correlation": 484, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1671.3703925084678
    }
  },
  {
    "ph": "f", "id": 484, "pid": 2, "tid": 0, "ts": 7253740902962.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907661.979, "dur": 3.840,
    "args": {
      "External id": 1660,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 841,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 841, "pid": 49625, "tid": 49625, "ts": 7253740907661.979,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903050.059, "dur": 5.050,
    "args": {
      "External id": 953,      "device": 2, "stream": 0,      "correlation": 485, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 485, "pid": 2, "tid": 0, "ts": 7253740903050.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903055.110, "dur": 5.412,
    "args": {
      "External id": 954,      "device": 2, "stream": 0,      "correlation": 486, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 486, "pid": 2, "tid": 0, "ts": 7253740903055.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907668.965, "dur": 3.323,
    "args": {
      "External id": 1661,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 842,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 842, "pid": 49625, "tid": 49625, "ts": 7253740907668.965,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740903060.523, "dur": 78.018,
    "args": {
      "External id": 955,      "device": 2, "stream": 0,      "correlation": 487, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 487, "pid": 2, "tid": 0, "ts": 7253740903060.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903138.542, "dur": 63.304,
    "args": {
      "External id": 956,      "device": 2, "stream": 0,      "correlation": 488, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 488, "pid": 2, "tid": 0, "ts": 7253740903138.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740903201.847, "dur": 76.776,
    "args": {
      "External id": 957,      "device": 2, "stream": 0,      "correlation": 489, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 489, "pid": 2, "tid": 0, "ts": 7253740903201.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903278.624, "dur": 66.832,
    "args": {
      "External id": 958,      "device": 2, "stream": 0,      "correlation": 490, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 490, "pid": 2, "tid": 0, "ts": 7253740903278.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740903345.457, "dur": 74.049,
    "args": {
      "External id": 960,      "device": 2, "stream": 0,      "correlation": 491, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 491, "pid": 2, "tid": 0, "ts": 7253740903345.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740903419.507, "dur": 72.766,
    "args": {
      "External id": 961,      "device": 2, "stream": 0,      "correlation": 492, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 492, "pid": 2, "tid": 0, "ts": 7253740903419.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903499.291, "dur": 44.421,
    "args": {
      "External id": 963,      "device": 2, "stream": 0,      "correlation": 493, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 493, "pid": 2, "tid": 0, "ts": 7253740903499.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740903543.713, "dur": 79.982,
    "args": {
      "External id": 966,      "device": 2, "stream": 0,      "correlation": 494, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1678.0781916157184
    }
  },
  {
    "ph": "f", "id": 494, "pid": 2, "tid": 0, "ts": 7253740903543.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907696.300, "dur": 4.014,
    "args": {
      "External id": 1663,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 843,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 843, "pid": 49625, "tid": 49625, "ts": 7253740907696.300,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903630.512, "dur": 5.211,
    "args": {
      "External id": 973,      "device": 2, "stream": 0,      "correlation": 495, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 495, "pid": 2, "tid": 0, "ts": 7253740903630.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903635.724, "dur": 5.451,
    "args": {
      "External id": 974,      "device": 2, "stream": 0,      "correlation": 496, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 496, "pid": 2, "tid": 0, "ts": 7253740903635.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740903641.176, "dur": 78.700,
    "args": {
      "External id": 975,      "device": 2, "stream": 0,      "correlation": 497, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 497, "pid": 2, "tid": 0, "ts": 7253740903641.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907707.777, "dur": 4.374,
    "args": {
      "External id": 1666,      "cid": 134, "correlation": 844, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 844, "pid": 49625, "tid": 49625, "ts": 7253740907707.777,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903719.877, "dur": 64.467,
    "args": {
      "External id": 976,      "device": 2, "stream": 0,      "correlation": 498, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 498, "pid": 2, "tid": 0, "ts": 7253740903719.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740903784.345, "dur": 77.377,
    "args": {
      "External id": 977,      "device": 2, "stream": 0,      "correlation": 499, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 499, "pid": 2, "tid": 0, "ts": 7253740903784.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740903861.723, "dur": 62.062,
    "args": {
      "External id": 978,      "device": 2, "stream": 0,      "correlation": 500, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 500, "pid": 2, "tid": 0, "ts": 7253740903861.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740903923.786, "dur": 70.881,
    "args": {
      "External id": 980,      "device": 2, "stream": 0,      "correlation": 501, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 501, "pid": 2, "tid": 0, "ts": 7253740903923.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740903994.668, "dur": 74.170,
    "args": {
      "External id": 981,      "device": 2, "stream": 0,      "correlation": 502, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 502, "pid": 2, "tid": 0, "ts": 7253740903994.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904076.176, "dur": 45.182,
    "args": {
      "External id": 983,      "device": 2, "stream": 0,      "correlation": 503, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 503, "pid": 2, "tid": 0, "ts": 7253740904076.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740904121.359, "dur": 78.139,
    "args": {
      "External id": 986,      "device": 2, "stream": 0,      "correlation": 504, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1717.6571282313796
    }
  },
  {
    "ph": "f", "id": 504, "pid": 2, "tid": 0, "ts": 7253740904121.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904206.595, "dur": 5.291,
    "args": {
      "External id": 993,      "device": 2, "stream": 0,      "correlation": 505, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 505, "pid": 2, "tid": 0, "ts": 7253740904206.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904211.887, "dur": 5.492,
    "args": {
      "External id": 994,      "device": 2, "stream": 0,      "correlation": 506, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 506, "pid": 2, "tid": 0, "ts": 7253740904211.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740904217.380, "dur": 74.891,
    "args": {
      "External id": 995,      "device": 2, "stream": 0,      "correlation": 507, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 507, "pid": 2, "tid": 0, "ts": 7253740904217.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904292.272, "dur": 58.934,
    "args": {
      "External id": 996,      "device": 2, "stream": 0,      "correlation": 508, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 508, "pid": 2, "tid": 0, "ts": 7253740904292.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740904351.207, "dur": 73.608,
    "args": {
      "External id": 997,      "device": 2, "stream": 0,      "correlation": 509, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 509, "pid": 2, "tid": 0, "ts": 7253740904351.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907754.559, "dur": 4.315,
    "args": {
      "External id": 1673,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 845,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 845, "pid": 49625, "tid": 49625, "ts": 7253740907754.559,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904424.816, "dur": 63.465,
    "args": {
      "External id": 998,      "device": 2, "stream": 0,      "correlation": 510, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 510, "pid": 2, "tid": 0, "ts": 7253740904424.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740904488.282, "dur": 70.481,
    "args": {
      "External id": 1000,      "device": 2, "stream": 0,      "correlation": 511, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 511, "pid": 2, "tid": 0, "ts": 7253740904488.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907763.744, "dur": 3.408,
    "args": {
      "External id": 1674,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 846,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 846, "pid": 49625, "tid": 49625, "ts": 7253740907763.744,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740904558.764, "dur": 74.891,
    "args": {
      "External id": 1001,      "device": 2, "stream": 0,      "correlation": 512, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 512, "pid": 2, "tid": 0, "ts": 7253740904558.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904641.474, "dur": 46.426,
    "args": {
      "External id": 1003,      "device": 2, "stream": 0,      "correlation": 513, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 513, "pid": 2, "tid": 0, "ts": 7253740904641.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740904687.901, "dur": 78.739,
    "args": {
      "External id": 1006,      "device": 2, "stream": 0,      "correlation": 514, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1704.5685547371095
    }
  },
  {
    "ph": "f", "id": 514, "pid": 2, "tid": 0, "ts": 7253740904687.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904773.297, "dur": 5.411,
    "args": {
      "External id": 1013,      "device": 2, "stream": 0,      "correlation": 515, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 515, "pid": 2, "tid": 0, "ts": 7253740904773.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907771.580, "dur": 4.533,
    "args": {
      "External id": 1675,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 847,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 847, "pid": 49625, "tid": 49625, "ts": 7253740907771.580,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904778.709, "dur": 5.812,
    "args": {
      "External id": 1014,      "device": 2, "stream": 0,      "correlation": 516, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 516, "pid": 2, "tid": 0, "ts": 7253740904778.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907779.787, "dur": 3.403,
    "args": {
      "External id": 1676,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 848,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 848, "pid": 49625, "tid": 49625, "ts": 7253740907779.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740904784.522, "dur": 79.502,
    "args": {
      "External id": 1015,      "device": 2, "stream": 0,      "correlation": 517, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 517, "pid": 2, "tid": 0, "ts": 7253740904784.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740904864.025, "dur": 62.142,
    "args": {
      "External id": 1016,      "device": 2, "stream": 0,      "correlation": 518, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 518, "pid": 2, "tid": 0, "ts": 7253740904864.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907787.324, "dur": 3.375,
    "args": {
      "External id": 1677,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 849,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 849, "pid": 49625, "tid": 49625, "ts": 7253740907787.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740904926.168, "dur": 76.575,
    "args": {
      "External id": 1017,      "device": 2, "stream": 0,      "correlation": 519, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 519, "pid": 2, "tid": 0, "ts": 7253740904926.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905002.744, "dur": 62.983,
    "args": {
      "External id": 1018,      "device": 2, "stream": 0,      "correlation": 520, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 520, "pid": 2, "tid": 0, "ts": 7253740905002.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907795.498, "dur": 4.217,
    "args": {
      "External id": 1678,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 850,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 850, "pid": 49625, "tid": 49625, "ts": 7253740907795.498,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740905065.728, "dur": 74.651,
    "args": {
      "External id": 1020,      "device": 2, "stream": 0,      "correlation": 521, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 521, "pid": 2, "tid": 0, "ts": 7253740905065.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907803.771, "dur": 3.573,
    "args": {
      "External id": 1680,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 851,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 851, "pid": 49625, "tid": 49625, "ts": 7253740907803.771,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740905140.380, "dur": 80.664,
    "args": {
      "External id": 1021,      "device": 2, "stream": 0,      "correlation": 522, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 522, "pid": 2, "tid": 0, "ts": 7253740905140.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905228.382, "dur": 46.145,
    "args": {
      "External id": 1023,      "device": 2, "stream": 0,      "correlation": 523, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 523, "pid": 2, "tid": 0, "ts": 7253740905228.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740905274.528, "dur": 76.294,
    "args": {
      "External id": 1026,      "device": 2, "stream": 0,      "correlation": 524, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1759.1942853397995
    }
  },
  {
    "ph": "f", "id": 524, "pid": 2, "tid": 0, "ts": 7253740905274.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907810.536, "dur": 3.048,
    "args": {
      "External id": 1681,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 852,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 852, "pid": 49625, "tid": 49625, "ts": 7253740907810.536,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905358.681, "dur": 5.492,
    "args": {
      "External id": 1033,      "device": 2, "stream": 0,      "correlation": 525, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 525, "pid": 2, "tid": 0, "ts": 7253740905358.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905364.174, "dur": 6.213,
    "args": {
      "External id": 1034,      "device": 2, "stream": 0,      "correlation": 526, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 526, "pid": 2, "tid": 0, "ts": 7253740905364.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740905370.388, "dur": 81.105,
    "args": {
      "External id": 1035,      "device": 2, "stream": 0,      "correlation": 527, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 527, "pid": 2, "tid": 0, "ts": 7253740905370.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905451.494, "dur": 64.147,
    "args": {
      "External id": 1036,      "device": 2, "stream": 0,      "correlation": 528, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 528, "pid": 2, "tid": 0, "ts": 7253740905451.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740905515.642, "dur": 75.973,
    "args": {
      "External id": 1037,      "device": 2, "stream": 0,      "correlation": 529, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 529, "pid": 2, "tid": 0, "ts": 7253740905515.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905591.616, "dur": 64.547,
    "args": {
      "External id": 1038,      "device": 2, "stream": 0,      "correlation": 530, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 530, "pid": 2, "tid": 0, "ts": 7253740905591.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740905656.164, "dur": 74.771,
    "args": {
      "External id": 1040,      "device": 2, "stream": 0,      "correlation": 531, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 531, "pid": 2, "tid": 0, "ts": 7253740905656.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907836.810, "dur": 3.775,
    "args": {
      "External id": 1683,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 853,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 853, "pid": 49625, "tid": 49625, "ts": 7253740907836.810,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740905730.936, "dur": 82.228,
    "args": {
      "External id": 1041,      "device": 2, "stream": 0,      "correlation": 532, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 532, "pid": 2, "tid": 0, "ts": 7253740905730.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905820.863, "dur": 46.145,
    "args": {
      "External id": 1043,      "device": 2, "stream": 0,      "correlation": 533, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 533, "pid": 2, "tid": 0, "ts": 7253740905820.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740905867.009, "dur": 77.938,
    "args": {
      "External id": 1046,      "device": 2, "stream": 0,      "correlation": 534, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1722.0868628029614
    }
  },
  {
    "ph": "f", "id": 534, "pid": 2, "tid": 0, "ts": 7253740905867.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905952.966, "dur": 5.532,
    "args": {
      "External id": 1053,      "device": 2, "stream": 0,      "correlation": 535, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 535, "pid": 2, "tid": 0, "ts": 7253740905952.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907848.435, "dur": 3.662,
    "args": {
      "External id": 1686,      "cid": 134, "correlation": 854, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 854, "pid": 49625, "tid": 49625, "ts": 7253740907848.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740905958.499, "dur": 5.692,
    "args": {
      "External id": 1054,      "device": 2, "stream": 0,      "correlation": 536, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 536, "pid": 2, "tid": 0, "ts": 7253740905958.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740905964.192, "dur": 83.711,
    "args": {
      "External id": 1055,      "device": 2, "stream": 0,      "correlation": 537, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 537, "pid": 2, "tid": 0, "ts": 7253740905964.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906047.904, "dur": 62.863,
    "args": {
      "External id": 1056,      "device": 2, "stream": 0,      "correlation": 538, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 538, "pid": 2, "tid": 0, "ts": 7253740906047.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740906110.768, "dur": 77.938,
    "args": {
      "External id": 1057,      "device": 2, "stream": 0,      "correlation": 539, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 539, "pid": 2, "tid": 0, "ts": 7253740906110.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906188.707, "dur": 65.751,
    "args": {
      "External id": 1058,      "device": 2, "stream": 0,      "correlation": 540, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 540, "pid": 2, "tid": 0, "ts": 7253740906188.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740906254.459, "dur": 76.013,
    "args": {
      "External id": 1060,      "device": 2, "stream": 0,      "correlation": 541, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 541, "pid": 2, "tid": 0, "ts": 7253740906254.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740906330.473, "dur": 83.030,
    "args": {
      "External id": 1061,      "device": 2, "stream": 0,      "correlation": 542, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 542, "pid": 2, "tid": 0, "ts": 7253740906330.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906420.520, "dur": 44.421,
    "args": {
      "External id": 1063,      "device": 2, "stream": 0,      "correlation": 543, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 543, "pid": 2, "tid": 0, "ts": 7253740906420.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740906464.942, "dur": 78.940,
    "args": {
      "External id": 1066,      "device": 2, "stream": 0,      "correlation": 544, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1700.2283730887625
    }
  },
  {
    "ph": "f", "id": 544, "pid": 2, "tid": 0, "ts": 7253740906464.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906550.819, "dur": 5.612,
    "args": {
      "External id": 1073,      "device": 2, "stream": 0,      "correlation": 545, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 545, "pid": 2, "tid": 0, "ts": 7253740906550.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906556.432, "dur": 5.973,
    "args": {
      "External id": 1074,      "device": 2, "stream": 0,      "correlation": 546, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 546, "pid": 2, "tid": 0, "ts": 7253740906556.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907891.711, "dur": 3.117,
    "args": {
      "External id": 1693,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 855,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 855, "pid": 49625, "tid": 49625, "ts": 7253740907891.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740906562.406, "dur": 83.671,
    "args": {
      "External id": 1075,      "device": 2, "stream": 0,      "correlation": 547, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 547, "pid": 2, "tid": 0, "ts": 7253740906562.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906646.078, "dur": 61.099,
    "args": {
      "External id": 1076,      "device": 2, "stream": 0,      "correlation": 548, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 548, "pid": 2, "tid": 0, "ts": 7253740906646.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907899.826, "dur": 3.043,
    "args": {
      "External id": 1694,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 856,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 856, "pid": 49625, "tid": 49625, "ts": 7253740907899.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740906707.178, "dur": 78.059,
    "args": {
      "External id": 1077,      "device": 2, "stream": 0,      "correlation": 549, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 549, "pid": 2, "tid": 0, "ts": 7253740906707.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907907.180, "dur": 4.775,
    "args": {
      "External id": 1695,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 857,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 857, "pid": 49625, "tid": 49625, "ts": 7253740907907.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740906785.238, "dur": 68.155,
    "args": {
      "External id": 1078,      "device": 2, "stream": 0,      "correlation": 550, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 550, "pid": 2, "tid": 0, "ts": 7253740906785.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740906853.394, "dur": 77.537,
    "args": {
      "External id": 1080,      "device": 2, "stream": 0,      "correlation": 551, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 551, "pid": 2, "tid": 0, "ts": 7253740906853.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907915.547, "dur": 3.434,
    "args": {
      "External id": 1696,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 858,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 858, "pid": 49625, "tid": 49625, "ts": 7253740907915.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740906930.932, "dur": 85.195,
    "args": {
      "External id": 1081,      "device": 2, "stream": 0,      "correlation": 552, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 552, "pid": 2, "tid": 0, "ts": 7253740906930.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907023.104, "dur": 44.862,
    "args": {
      "External id": 1083,      "device": 2, "stream": 0,      "correlation": 553, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 553, "pid": 2, "tid": 0, "ts": 7253740907023.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740907067.967, "dur": 77.497,
    "args": {
      "External id": 1086,      "device": 2, "stream": 0,      "correlation": 554, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1731.8863454540763
    }
  },
  {
    "ph": "f", "id": 554, "pid": 2, "tid": 0, "ts": 7253740907067.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907922.981, "dur": 3.063,
    "args": {
      "External id": 1697,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 859,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 859, "pid": 49625, "tid": 49625, "ts": 7253740907922.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907152.641, "dur": 5.452,
    "args": {
      "External id": 1093,      "device": 2, "stream": 0,      "correlation": 555, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 555, "pid": 2, "tid": 0, "ts": 7253740907152.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907158.094, "dur": 5.692,
    "args": {
      "External id": 1094,      "device": 2, "stream": 0,      "correlation": 556, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 556, "pid": 2, "tid": 0, "ts": 7253740907158.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740907163.787, "dur": 85.555,
    "args": {
      "External id": 1095,      "device": 2, "stream": 0,      "correlation": 557, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 557, "pid": 2, "tid": 0, "ts": 7253740907163.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907930.762, "dur": 6.252,
    "args": {
      "External id": 1698,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 860,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 860, "pid": 49625, "tid": 49625, "ts": 7253740907930.762,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907249.343, "dur": 64.307,
    "args": {
      "External id": 1096,      "device": 2, "stream": 0,      "correlation": 558, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 558, "pid": 2, "tid": 0, "ts": 7253740907249.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907941.027, "dur": 3.746,
    "args": {
      "External id": 1700,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 861,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 861, "pid": 49625, "tid": 49625, "ts": 7253740907941.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907947.956, "dur": 3.361,
    "args": {
      "External id": 1701,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 862,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 862, "pid": 49625, "tid": 49625, "ts": 7253740907947.956,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740907974.957, "dur": 4.561,
    "args": {
      "External id": 1703,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 863,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 863, "pid": 49625, "tid": 49625, "ts": 7253740907974.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740907987.259, "dur": 3.916,
    "args": {
      "External id": 1706,      "cid": 134, "correlation": 864, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 864, "pid": 49625, "tid": 49625, "ts": 7253740907987.259,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908032.351, "dur": 4.307,
    "args": {
      "External id": 1713,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 865,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 865, "pid": 49625, "tid": 49625, "ts": 7253740908032.351,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908041.542, "dur": 4.069,
    "args": {
      "External id": 1714,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 866,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 866, "pid": 49625, "tid": 49625, "ts": 7253740908041.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908049.760, "dur": 3.874,
    "args": {
      "External id": 1715,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 867,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 867, "pid": 49625, "tid": 49625, "ts": 7253740908049.760,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908056.964, "dur": 3.403,
    "args": {
      "External id": 1716,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 868,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 868, "pid": 49625, "tid": 49625, "ts": 7253740908056.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908064.418, "dur": 5.068,
    "args": {
      "External id": 1717,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 869,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 869, "pid": 49625, "tid": 49625, "ts": 7253740908064.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908074.284, "dur": 3.563,
    "args": {
      "External id": 1718,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 870,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 870, "pid": 49625, "tid": 49625, "ts": 7253740908074.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908081.630, "dur": 3.686,
    "args": {
      "External id": 1720,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 871,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 871, "pid": 49625, "tid": 49625, "ts": 7253740908081.630,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908088.581, "dur": 4.645,
    "args": {
      "External id": 1721,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 872,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 872, "pid": 49625, "tid": 49625, "ts": 7253740908088.581,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908117.500, "dur": 4.121,
    "args": {
      "External id": 1723,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 873,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 873, "pid": 49625, "tid": 49625, "ts": 7253740908117.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908129.613, "dur": 4.719,
    "args": {
      "External id": 1726,      "cid": 134, "correlation": 874, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 874, "pid": 49625, "tid": 49625, "ts": 7253740908129.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908175.206, "dur": 3.860,
    "args": {
      "External id": 1733,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 875,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 875, "pid": 49625, "tid": 49625, "ts": 7253740908175.206,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908183.841, "dur": 3.510,
    "args": {
      "External id": 1734,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 876,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 876, "pid": 49625, "tid": 49625, "ts": 7253740908183.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908191.513, "dur": 3.385,
    "args": {
      "External id": 1735,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 877,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 877, "pid": 49625, "tid": 49625, "ts": 7253740908191.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908198.267, "dur": 3.695,
    "args": {
      "External id": 1736,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 878,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 878, "pid": 49625, "tid": 49625, "ts": 7253740908198.267,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908206.015, "dur": 4.045,
    "args": {
      "External id": 1737,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 879,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 879, "pid": 49625, "tid": 49625, "ts": 7253740908206.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908214.811, "dur": 3.313,
    "args": {
      "External id": 1738,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 880,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 880, "pid": 49625, "tid": 49625, "ts": 7253740908214.811,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908221.897, "dur": 4.228,
    "args": {
      "External id": 1740,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 881,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 881, "pid": 49625, "tid": 49625, "ts": 7253740908221.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908235.623, "dur": 4.364,
    "args": {
      "External id": 1741,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 882,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 882, "pid": 49625, "tid": 49625, "ts": 7253740908235.623,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908265.340, "dur": 5.227,
    "args": {
      "External id": 1743,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 883,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 883, "pid": 49625, "tid": 49625, "ts": 7253740908265.340,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908278.531, "dur": 6.018,
    "args": {
      "External id": 1746,      "cid": 134, "correlation": 884, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 884, "pid": 49625, "tid": 49625, "ts": 7253740908278.531,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908327.197, "dur": 4.124,
    "args": {
      "External id": 1753,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 885,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 885, "pid": 49625, "tid": 49625, "ts": 7253740908327.197,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908337.219, "dur": 4.893,
    "args": {
      "External id": 1754,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 886,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 886, "pid": 49625, "tid": 49625, "ts": 7253740908337.219,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908346.530, "dur": 4.200,
    "args": {
      "External id": 1755,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 887,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 887, "pid": 49625, "tid": 49625, "ts": 7253740908346.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908354.400, "dur": 4.316,
    "args": {
      "External id": 1756,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 888,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 888, "pid": 49625, "tid": 49625, "ts": 7253740908354.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908363.234, "dur": 4.630,
    "args": {
      "External id": 1757,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 889,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 889, "pid": 49625, "tid": 49625, "ts": 7253740908363.234,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908372.764, "dur": 3.420,
    "args": {
      "External id": 1758,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 890,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 890, "pid": 49625, "tid": 49625, "ts": 7253740908372.764,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908380.277, "dur": 4.884,
    "args": {
      "External id": 1760,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 891,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 891, "pid": 49625, "tid": 49625, "ts": 7253740908380.277,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908388.863, "dur": 4.970,
    "args": {
      "External id": 1761,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 892,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 892, "pid": 49625, "tid": 49625, "ts": 7253740908388.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908419.652, "dur": 4.619,
    "args": {
      "External id": 1763,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 893,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 893, "pid": 49625, "tid": 49625, "ts": 7253740908419.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908432.154, "dur": 4.666,
    "args": {
      "External id": 1766,      "cid": 134, "correlation": 894, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 894, "pid": 49625, "tid": 49625, "ts": 7253740908432.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908477.767, "dur": 5.019,
    "args": {
      "External id": 1773,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 895,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 895, "pid": 49625, "tid": 49625, "ts": 7253740908477.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908487.661, "dur": 4.203,
    "args": {
      "External id": 1774,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 896,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 896, "pid": 49625, "tid": 49625, "ts": 7253740908487.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908496.146, "dur": 4.535,
    "args": {
      "External id": 1775,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 897,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 897, "pid": 49625, "tid": 49625, "ts": 7253740908496.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908504.030, "dur": 3.727,
    "args": {
      "External id": 1776,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 898,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 898, "pid": 49625, "tid": 49625, "ts": 7253740908504.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908511.742, "dur": 3.359,
    "args": {
      "External id": 1777,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 899,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 899, "pid": 49625, "tid": 49625, "ts": 7253740908511.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908519.830, "dur": 3.925,
    "args": {
      "External id": 1778,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 900,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 900, "pid": 49625, "tid": 49625, "ts": 7253740908519.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908527.703, "dur": 4.217,
    "args": {
      "External id": 1780,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 901,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 901, "pid": 49625, "tid": 49625, "ts": 7253740908527.703,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908535.040, "dur": 3.922,
    "args": {
      "External id": 1781,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 902,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 902, "pid": 49625, "tid": 49625, "ts": 7253740908535.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908562.627, "dur": 4.402,
    "args": {
      "External id": 1783,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 903,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 903, "pid": 49625, "tid": 49625, "ts": 7253740908562.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908574.557, "dur": 4.569,
    "args": {
      "External id": 1786,      "cid": 134, "correlation": 904, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 904, "pid": 49625, "tid": 49625, "ts": 7253740908574.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908618.964, "dur": 4.513,
    "args": {
      "External id": 1793,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 905,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 905, "pid": 49625, "tid": 49625, "ts": 7253740908618.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908628.324, "dur": 3.677,
    "args": {
      "External id": 1794,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 906,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 906, "pid": 49625, "tid": 49625, "ts": 7253740908628.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908636.268, "dur": 4.626,
    "args": {
      "External id": 1795,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 907,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 907, "pid": 49625, "tid": 49625, "ts": 7253740908636.268,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908644.274, "dur": 3.616,
    "args": {
      "External id": 1796,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 908,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 908, "pid": 49625, "tid": 49625, "ts": 7253740908644.274,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908652.036, "dur": 3.252,
    "args": {
      "External id": 1797,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 909,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 909, "pid": 49625, "tid": 49625, "ts": 7253740908652.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908659.966, "dur": 3.824,
    "args": {
      "External id": 1798,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 910,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 910, "pid": 49625, "tid": 49625, "ts": 7253740908659.966,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908667.752, "dur": 3.798,
    "args": {
      "External id": 1800,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 911,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 911, "pid": 49625, "tid": 49625, "ts": 7253740908667.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908674.674, "dur": 3.384,
    "args": {
      "External id": 1801,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 912,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 912, "pid": 49625, "tid": 49625, "ts": 7253740908674.674,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908701.699, "dur": 4.423,
    "args": {
      "External id": 1803,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 913,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 913, "pid": 49625, "tid": 49625, "ts": 7253740908701.699,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908713.471, "dur": 4.177,
    "args": {
      "External id": 1806,      "cid": 134, "correlation": 914, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 914, "pid": 49625, "tid": 49625, "ts": 7253740908713.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908757.312, "dur": 4.267,
    "args": {
      "External id": 1813,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 915,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 915, "pid": 49625, "tid": 49625, "ts": 7253740908757.312,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908766.608, "dur": 3.526,
    "args": {
      "External id": 1814,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 916,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 916, "pid": 49625, "tid": 49625, "ts": 7253740908766.608,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908774.467, "dur": 3.535,
    "args": {
      "External id": 1815,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 917,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 917, "pid": 49625, "tid": 49625, "ts": 7253740908774.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908781.387, "dur": 3.701,
    "args": {
      "External id": 1816,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 918,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 918, "pid": 49625, "tid": 49625, "ts": 7253740908781.387,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908789.202, "dur": 4.113,
    "args": {
      "External id": 1817,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 919,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 919, "pid": 49625, "tid": 49625, "ts": 7253740908789.202,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908798.045, "dur": 3.558,
    "args": {
      "External id": 1818,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 920,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 920, "pid": 49625, "tid": 49625, "ts": 7253740908798.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908805.379, "dur": 3.536,
    "args": {
      "External id": 1820,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 921,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 921, "pid": 49625, "tid": 49625, "ts": 7253740908805.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908812.138, "dur": 3.314,
    "args": {
      "External id": 1821,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 922,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 922, "pid": 49625, "tid": 49625, "ts": 7253740908812.138,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908854.766, "dur": 4.597,
    "args": {
      "External id": 1823,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 923,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 923, "pid": 49625, "tid": 49625, "ts": 7253740908854.766,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740908869.975, "dur": 6.655,
    "args": {
      "External id": 1826,      "cid": 134, "correlation": 924, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 924, "pid": 49625, "tid": 49625, "ts": 7253740908869.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908924.027, "dur": 5.621,
    "args": {
      "External id": 1833,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 925,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 925, "pid": 49625, "tid": 49625, "ts": 7253740908924.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908935.692, "dur": 3.750,
    "args": {
      "External id": 1834,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 926,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 926, "pid": 49625, "tid": 49625, "ts": 7253740908935.692,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908944.138, "dur": 3.055,
    "args": {
      "External id": 1835,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 927,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 927, "pid": 49625, "tid": 49625, "ts": 7253740908944.138,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908950.918, "dur": 4.584,
    "args": {
      "External id": 1836,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 928,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 928, "pid": 49625, "tid": 49625, "ts": 7253740908950.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908960.591, "dur": 4.948,
    "args": {
      "External id": 1837,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 929,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 929, "pid": 49625, "tid": 49625, "ts": 7253740908960.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908970.742, "dur": 3.793,
    "args": {
      "External id": 1838,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 930,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 930, "pid": 49625, "tid": 49625, "ts": 7253740908970.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908978.746, "dur": 5.134,
    "args": {
      "External id": 1840,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 931,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 931, "pid": 49625, "tid": 49625, "ts": 7253740908978.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740908987.107, "dur": 3.538,
    "args": {
      "External id": 1841,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 932,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 932, "pid": 49625, "tid": 49625, "ts": 7253740908987.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909017.048, "dur": 4.643,
    "args": {
      "External id": 1843,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 933,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 933, "pid": 49625, "tid": 49625, "ts": 7253740909017.048,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909029.765, "dur": 4.714,
    "args": {
      "External id": 1846,      "cid": 134, "correlation": 934, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 934, "pid": 49625, "tid": 49625, "ts": 7253740909029.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909077.395, "dur": 5.180,
    "args": {
      "External id": 1853,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 935,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 935, "pid": 49625, "tid": 49625, "ts": 7253740909077.395,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909087.711, "dur": 3.644,
    "args": {
      "External id": 1854,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 936,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 936, "pid": 49625, "tid": 49625, "ts": 7253740909087.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909095.719, "dur": 3.492,
    "args": {
      "External id": 1855,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 937,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 937, "pid": 49625, "tid": 49625, "ts": 7253740909095.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909102.953, "dur": 3.247,
    "args": {
      "External id": 1856,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 938,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 938, "pid": 49625, "tid": 49625, "ts": 7253740909102.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909110.501, "dur": 3.658,
    "args": {
      "External id": 1857,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 939,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 939, "pid": 49625, "tid": 49625, "ts": 7253740909110.501,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909119.055, "dur": 3.973,
    "args": {
      "External id": 1858,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 940,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 940, "pid": 49625, "tid": 49625, "ts": 7253740909119.055,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909127.070, "dur": 3.472,
    "args": {
      "External id": 1860,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 941,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 941, "pid": 49625, "tid": 49625, "ts": 7253740909127.070,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909133.722, "dur": 3.833,
    "args": {
      "External id": 1861,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 942,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 942, "pid": 49625, "tid": 49625, "ts": 7253740909133.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909162.323, "dur": 4.153,
    "args": {
      "External id": 1863,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 943,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 943, "pid": 49625, "tid": 49625, "ts": 7253740909162.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909174.143, "dur": 4.409,
    "args": {
      "External id": 1866,      "cid": 134, "correlation": 944, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 944, "pid": 49625, "tid": 49625, "ts": 7253740909174.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909219.531, "dur": 6.527,
    "args": {
      "External id": 1873,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 945,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 945, "pid": 49625, "tid": 49625, "ts": 7253740909219.531,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909231.037, "dur": 4.045,
    "args": {
      "External id": 1874,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 946,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 946, "pid": 49625, "tid": 49625, "ts": 7253740909231.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909239.430, "dur": 3.792,
    "args": {
      "External id": 1875,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 947,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 947, "pid": 49625, "tid": 49625, "ts": 7253740909239.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909246.534, "dur": 3.197,
    "args": {
      "External id": 1876,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 948,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 948, "pid": 49625, "tid": 49625, "ts": 7253740909246.534,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909253.832, "dur": 3.822,
    "args": {
      "External id": 1877,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 949,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 949, "pid": 49625, "tid": 49625, "ts": 7253740909253.832,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909262.341, "dur": 4.098,
    "args": {
      "External id": 1878,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 950,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 950, "pid": 49625, "tid": 49625, "ts": 7253740909262.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909270.293, "dur": 3.636,
    "args": {
      "External id": 1880,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 951,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 951, "pid": 49625, "tid": 49625, "ts": 7253740909270.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909277.026, "dur": 3.574,
    "args": {
      "External id": 1881,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 952,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 952, "pid": 49625, "tid": 49625, "ts": 7253740909277.026,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909305.226, "dur": 4.039,
    "args": {
      "External id": 1883,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 953,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 953, "pid": 49625, "tid": 49625, "ts": 7253740909305.226,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909317.074, "dur": 4.037,
    "args": {
      "External id": 1886,      "cid": 134, "correlation": 954, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 954, "pid": 49625, "tid": 49625, "ts": 7253740909317.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909361.855, "dur": 4.359,
    "args": {
      "External id": 1893,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 955,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 955, "pid": 49625, "tid": 49625, "ts": 7253740909361.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909371.157, "dur": 3.729,
    "args": {
      "External id": 1894,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 956,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 956, "pid": 49625, "tid": 49625, "ts": 7253740909371.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909379.229, "dur": 4.137,
    "args": {
      "External id": 1895,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 957,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 957, "pid": 49625, "tid": 49625, "ts": 7253740909379.229,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909386.896, "dur": 3.806,
    "args": {
      "External id": 1896,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 958,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 958, "pid": 49625, "tid": 49625, "ts": 7253740909386.896,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909394.731, "dur": 3.899,
    "args": {
      "External id": 1897,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 959,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 959, "pid": 49625, "tid": 49625, "ts": 7253740909394.731,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909403.393, "dur": 3.296,
    "args": {
      "External id": 1898,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 960,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 960, "pid": 49625, "tid": 49625, "ts": 7253740909403.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909410.536, "dur": 3.834,
    "args": {
      "External id": 1900,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 961,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 961, "pid": 49625, "tid": 49625, "ts": 7253740909410.536,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909421.769, "dur": 5.591,
    "args": {
      "External id": 1901,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 962,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 962, "pid": 49625, "tid": 49625, "ts": 7253740909421.769,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909453.622, "dur": 5.080,
    "args": {
      "External id": 1903,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 963,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 963, "pid": 49625, "tid": 49625, "ts": 7253740909453.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909466.753, "dur": 4.674,
    "args": {
      "External id": 1906,      "cid": 134, "correlation": 964, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 964, "pid": 49625, "tid": 49625, "ts": 7253740909466.753,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909514.055, "dur": 4.249,
    "args": {
      "External id": 1913,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 965,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 965, "pid": 49625, "tid": 49625, "ts": 7253740909514.055,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909523.342, "dur": 3.708,
    "args": {
      "External id": 1914,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 966,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 966, "pid": 49625, "tid": 49625, "ts": 7253740909523.342,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909531.205, "dur": 3.529,
    "args": {
      "External id": 1915,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 967,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 967, "pid": 49625, "tid": 49625, "ts": 7253740909531.205,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909537.979, "dur": 3.521,
    "args": {
      "External id": 1916,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 968,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 968, "pid": 49625, "tid": 49625, "ts": 7253740909537.979,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909546.348, "dur": 3.877,
    "args": {
      "External id": 1917,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 969,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 969, "pid": 49625, "tid": 49625, "ts": 7253740909546.348,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909555.121, "dur": 4.633,
    "args": {
      "External id": 1918,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 970,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 970, "pid": 49625, "tid": 49625, "ts": 7253740909555.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909563.721, "dur": 3.882,
    "args": {
      "External id": 1920,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 971,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 971, "pid": 49625, "tid": 49625, "ts": 7253740909563.721,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909570.825, "dur": 3.423,
    "args": {
      "External id": 1921,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 972,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 972, "pid": 49625, "tid": 49625, "ts": 7253740909570.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909599.693, "dur": 4.209,
    "args": {
      "External id": 1923,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 973,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 973, "pid": 49625, "tid": 49625, "ts": 7253740909599.693,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909611.478, "dur": 4.216,
    "args": {
      "External id": 1926,      "cid": 134, "correlation": 974, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 974, "pid": 49625, "tid": 49625, "ts": 7253740909611.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909656.627, "dur": 3.416,
    "args": {
      "External id": 1933,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 975,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 975, "pid": 49625, "tid": 49625, "ts": 7253740909656.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909664.915, "dur": 4.019,
    "args": {
      "External id": 1934,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 976,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 976, "pid": 49625, "tid": 49625, "ts": 7253740909664.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909673.431, "dur": 3.507,
    "args": {
      "External id": 1935,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 977,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 977, "pid": 49625, "tid": 49625, "ts": 7253740909673.431,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909680.223, "dur": 3.317,
    "args": {
      "External id": 1936,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 978,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 978, "pid": 49625, "tid": 49625, "ts": 7253740909680.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909687.671, "dur": 3.615,
    "args": {
      "External id": 1937,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 979,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 979, "pid": 49625, "tid": 49625, "ts": 7253740909687.671,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909696.071, "dur": 3.048,
    "args": {
      "External id": 1938,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 980,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 980, "pid": 49625, "tid": 49625, "ts": 7253740909696.071,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909703.027, "dur": 3.536,
    "args": {
      "External id": 1940,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 981,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 981, "pid": 49625, "tid": 49625, "ts": 7253740909703.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909709.808, "dur": 3.846,
    "args": {
      "External id": 1941,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 982,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 982, "pid": 49625, "tid": 49625, "ts": 7253740909709.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909737.560, "dur": 3.988,
    "args": {
      "External id": 1943,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 983,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 983, "pid": 49625, "tid": 49625, "ts": 7253740909737.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909749.308, "dur": 4.032,
    "args": {
      "External id": 1946,      "cid": 134, "correlation": 984, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 984, "pid": 49625, "tid": 49625, "ts": 7253740909749.308,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909793.504, "dur": 3.833,
    "args": {
      "External id": 1953,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 985,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 985, "pid": 49625, "tid": 49625, "ts": 7253740909793.504,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909802.042, "dur": 4.264,
    "args": {
      "External id": 1954,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 986,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 986, "pid": 49625, "tid": 49625, "ts": 7253740909802.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909810.570, "dur": 3.724,
    "args": {
      "External id": 1955,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 987,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 987, "pid": 49625, "tid": 49625, "ts": 7253740909810.570,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909817.607, "dur": 3.781,
    "args": {
      "External id": 1956,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 988,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 988, "pid": 49625, "tid": 49625, "ts": 7253740909817.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909825.475, "dur": 4.331,
    "args": {
      "External id": 1957,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 989,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 989, "pid": 49625, "tid": 49625, "ts": 7253740909825.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909834.544, "dur": 3.122,
    "args": {
      "External id": 1958,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 990,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 990, "pid": 49625, "tid": 49625, "ts": 7253740909834.544,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909841.558, "dur": 3.769,
    "args": {
      "External id": 1960,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 991,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 991, "pid": 49625, "tid": 49625, "ts": 7253740909841.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909848.492, "dur": 3.774,
    "args": {
      "External id": 1961,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 992,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 992, "pid": 49625, "tid": 49625, "ts": 7253740909848.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909876.525, "dur": 3.882,
    "args": {
      "External id": 1963,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 993,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 993, "pid": 49625, "tid": 49625, "ts": 7253740909876.525,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740909887.963, "dur": 4.290,
    "args": {
      "External id": 1966,      "cid": 134, "correlation": 994, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 994, "pid": 49625, "tid": 49625, "ts": 7253740909887.963,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909932.828, "dur": 3.620,
    "args": {
      "External id": 1973,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 995,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 995, "pid": 49625, "tid": 49625, "ts": 7253740909932.828,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909941.169, "dur": 4.404,
    "args": {
      "External id": 1974,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 996,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 996, "pid": 49625, "tid": 49625, "ts": 7253740909941.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909949.885, "dur": 3.796,
    "args": {
      "External id": 1975,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 997,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 997, "pid": 49625, "tid": 49625, "ts": 7253740909949.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909957.031, "dur": 3.813,
    "args": {
      "External id": 1976,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 998,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 998, "pid": 49625, "tid": 49625, "ts": 7253740909957.031,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909964.904, "dur": 4.360,
    "args": {
      "External id": 1977,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 999,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 999, "pid": 49625, "tid": 49625, "ts": 7253740909964.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909974.118, "dur": 3.972,
    "args": {
      "External id": 1978,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1000,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1000, "pid": 49625, "tid": 49625, "ts": 7253740909974.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909981.975, "dur": 3.577,
    "args": {
      "External id": 1980,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1001,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1001, "pid": 49625, "tid": 49625, "ts": 7253740909981.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740909988.815, "dur": 3.603,
    "args": {
      "External id": 1981,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1002,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1002, "pid": 49625, "tid": 49625, "ts": 7253740909988.815,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910017.574, "dur": 4.135,
    "args": {
      "External id": 1983,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1003,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1003, "pid": 49625, "tid": 49625, "ts": 7253740910017.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910030.015, "dur": 5.159,
    "args": {
      "External id": 1986,      "cid": 134, "correlation": 1004, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1004, "pid": 49625, "tid": 49625, "ts": 7253740910030.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910095.741, "dur": 5.594,
    "args": {
      "External id": 1993,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1005,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1005, "pid": 49625, "tid": 49625, "ts": 7253740910095.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910108.239, "dur": 4.616,
    "args": {
      "External id": 1994,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1006,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1006, "pid": 49625, "tid": 49625, "ts": 7253740910108.239,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910117.879, "dur": 3.824,
    "args": {
      "External id": 1995,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1007,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1007, "pid": 49625, "tid": 49625, "ts": 7253740910117.879,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910125.356, "dur": 3.591,
    "args": {
      "External id": 1996,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1008,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1008, "pid": 49625, "tid": 49625, "ts": 7253740910125.356,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910136.999, "dur": 4.558,
    "args": {
      "External id": 1997,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1009,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1009, "pid": 49625, "tid": 49625, "ts": 7253740910136.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910147.460, "dur": 5.075,
    "args": {
      "External id": 1998,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1010,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1010, "pid": 49625, "tid": 49625, "ts": 7253740910147.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910156.909, "dur": 14.391,
    "args": {
      "External id": 2000,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1011,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1011, "pid": 49625, "tid": 49625, "ts": 7253740910156.909,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910174.742, "dur": 4.154,
    "args": {
      "External id": 2001,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1012,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1012, "pid": 49625, "tid": 49625, "ts": 7253740910174.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910208.534, "dur": 4.481,
    "args": {
      "External id": 2003,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1013,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1013, "pid": 49625, "tid": 49625, "ts": 7253740910208.534,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910222.563, "dur": 4.370,
    "args": {
      "External id": 2006,      "cid": 134, "correlation": 1014, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1014, "pid": 49625, "tid": 49625, "ts": 7253740910222.563,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910272.802, "dur": 4.238,
    "args": {
      "External id": 2013,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1015,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1015, "pid": 49625, "tid": 49625, "ts": 7253740910272.802,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910282.444, "dur": 3.732,
    "args": {
      "External id": 2014,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1016,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1016, "pid": 49625, "tid": 49625, "ts": 7253740910282.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910290.608, "dur": 3.981,
    "args": {
      "External id": 2015,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1017,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1017, "pid": 49625, "tid": 49625, "ts": 7253740910290.608,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910298.161, "dur": 4.178,
    "args": {
      "External id": 2016,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1018,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1018, "pid": 49625, "tid": 49625, "ts": 7253740910298.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910306.690, "dur": 4.007,
    "args": {
      "External id": 2017,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1019,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1019, "pid": 49625, "tid": 49625, "ts": 7253740910306.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910315.747, "dur": 3.875,
    "args": {
      "External id": 2018,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1020,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1020, "pid": 49625, "tid": 49625, "ts": 7253740910315.747,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910323.615, "dur": 3.631,
    "args": {
      "External id": 2020,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1021,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1021, "pid": 49625, "tid": 49625, "ts": 7253740910323.615,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910330.309, "dur": 3.603,
    "args": {
      "External id": 2021,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1022,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1022, "pid": 49625, "tid": 49625, "ts": 7253740910330.309,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910360.606, "dur": 3.915,
    "args": {
      "External id": 2023,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1023,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1023, "pid": 49625, "tid": 49625, "ts": 7253740910360.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910372.275, "dur": 4.593,
    "args": {
      "External id": 2026,      "cid": 134, "correlation": 1024, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1024, "pid": 49625, "tid": 49625, "ts": 7253740910372.275,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910420.121, "dur": 4.840,
    "args": {
      "External id": 2033,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1025,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1025, "pid": 49625, "tid": 49625, "ts": 7253740910420.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910430.244, "dur": 3.473,
    "args": {
      "External id": 2034,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1026,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1026, "pid": 49625, "tid": 49625, "ts": 7253740910430.244,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910438.115, "dur": 3.654,
    "args": {
      "External id": 2035,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1027,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1027, "pid": 49625, "tid": 49625, "ts": 7253740910438.115,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910446.988, "dur": 4.003,
    "args": {
      "External id": 2036,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1028,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1028, "pid": 49625, "tid": 49625, "ts": 7253740910446.988,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910455.346, "dur": 3.937,
    "args": {
      "External id": 2037,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1029,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1029, "pid": 49625, "tid": 49625, "ts": 7253740910455.346,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910464.468, "dur": 3.391,
    "args": {
      "External id": 2038,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1030,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1030, "pid": 49625, "tid": 49625, "ts": 7253740910464.468,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910471.800, "dur": 3.755,
    "args": {
      "External id": 2040,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1031,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1031, "pid": 49625, "tid": 49625, "ts": 7253740910471.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910478.668, "dur": 3.341,
    "args": {
      "External id": 2041,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1032,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1032, "pid": 49625, "tid": 49625, "ts": 7253740910478.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910507.911, "dur": 3.326,
    "args": {
      "External id": 2043,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1033,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1033, "pid": 49625, "tid": 49625, "ts": 7253740910507.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910518.897, "dur": 4.415,
    "args": {
      "External id": 2046,      "cid": 134, "correlation": 1034, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1034, "pid": 49625, "tid": 49625, "ts": 7253740910518.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910576.092, "dur": 4.458,
    "args": {
      "External id": 2053,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1035,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1035, "pid": 49625, "tid": 49625, "ts": 7253740910576.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910585.982, "dur": 4.285,
    "args": {
      "External id": 2054,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1036,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1036, "pid": 49625, "tid": 49625, "ts": 7253740910585.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910594.748, "dur": 3.365,
    "args": {
      "External id": 2055,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1037,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1037, "pid": 49625, "tid": 49625, "ts": 7253740910594.748,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910606.084, "dur": 4.232,
    "args": {
      "External id": 2056,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1038,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1038, "pid": 49625, "tid": 49625, "ts": 7253740910606.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910615.514, "dur": 7.289,
    "args": {
      "External id": 2057,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1039,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1039, "pid": 49625, "tid": 49625, "ts": 7253740910615.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910628.539, "dur": 4.227,
    "args": {
      "External id": 2058,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1040,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1040, "pid": 49625, "tid": 49625, "ts": 7253740910628.539,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910637.014, "dur": 3.914,
    "args": {
      "External id": 2060,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1041,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1041, "pid": 49625, "tid": 49625, "ts": 7253740910637.014,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910644.337, "dur": 4.158,
    "args": {
      "External id": 2061,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1042,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1042, "pid": 49625, "tid": 49625, "ts": 7253740910644.337,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910676.310, "dur": 4.460,
    "args": {
      "External id": 2063,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1043,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1043, "pid": 49625, "tid": 49625, "ts": 7253740910676.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910689.217, "dur": 4.786,
    "args": {
      "External id": 2066,      "cid": 134, "correlation": 1044, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1044, "pid": 49625, "tid": 49625, "ts": 7253740910689.217,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910740.944, "dur": 4.544,
    "args": {
      "External id": 2073,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1045,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1045, "pid": 49625, "tid": 49625, "ts": 7253740910740.944,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910750.982, "dur": 3.642,
    "args": {
      "External id": 2074,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1046,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1046, "pid": 49625, "tid": 49625, "ts": 7253740910750.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910759.406, "dur": 5.199,
    "args": {
      "External id": 2075,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1047,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1047, "pid": 49625, "tid": 49625, "ts": 7253740910759.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910768.478, "dur": 3.482,
    "args": {
      "External id": 2076,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1048,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1048, "pid": 49625, "tid": 49625, "ts": 7253740910768.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910776.495, "dur": 4.397,
    "args": {
      "External id": 2077,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1049,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1049, "pid": 49625, "tid": 49625, "ts": 7253740910776.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910786.054, "dur": 3.802,
    "args": {
      "External id": 2078,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1050,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1050, "pid": 49625, "tid": 49625, "ts": 7253740910786.054,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910794.143, "dur": 4.471,
    "args": {
      "External id": 2080,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1051,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1051, "pid": 49625, "tid": 49625, "ts": 7253740910794.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910801.982, "dur": 3.962,
    "args": {
      "External id": 2081,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1052,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1052, "pid": 49625, "tid": 49625, "ts": 7253740910801.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910832.873, "dur": 5.220,
    "args": {
      "External id": 2083,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1053,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1053, "pid": 49625, "tid": 49625, "ts": 7253740910832.873,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910846.473, "dur": 4.955,
    "args": {
      "External id": 2086,      "cid": 134, "correlation": 1054, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1054, "pid": 49625, "tid": 49625, "ts": 7253740910846.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910897.360, "dur": 3.598,
    "args": {
      "External id": 2093,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1055,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1055, "pid": 49625, "tid": 49625, "ts": 7253740910897.360,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910906.080, "dur": 4.448,
    "args": {
      "External id": 2094,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1056,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1056, "pid": 49625, "tid": 49625, "ts": 7253740910906.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910915.673, "dur": 3.863,
    "args": {
      "External id": 2095,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1057,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1057, "pid": 49625, "tid": 49625, "ts": 7253740910915.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910923.180, "dur": 3.814,
    "args": {
      "External id": 2096,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1058,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1058, "pid": 49625, "tid": 49625, "ts": 7253740910923.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910931.420, "dur": 4.006,
    "args": {
      "External id": 2097,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1059,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1059, "pid": 49625, "tid": 49625, "ts": 7253740910931.420,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910940.668, "dur": 3.875,
    "args": {
      "External id": 2098,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1060,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1060, "pid": 49625, "tid": 49625, "ts": 7253740910940.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910948.823, "dur": 3.835,
    "args": {
      "External id": 2100,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1061,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1061, "pid": 49625, "tid": 49625, "ts": 7253740910948.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910955.997, "dur": 3.431,
    "args": {
      "External id": 2101,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1062,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1062, "pid": 49625, "tid": 49625, "ts": 7253740910955.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740910986.564, "dur": 4.360,
    "args": {
      "External id": 2103,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1063,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1063, "pid": 49625, "tid": 49625, "ts": 7253740910986.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740910999.114, "dur": 4.322,
    "args": {
      "External id": 2106,      "cid": 134, "correlation": 1064, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1064, "pid": 49625, "tid": 49625, "ts": 7253740910999.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911048.215, "dur": 3.889,
    "args": {
      "External id": 2113,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1065,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1065, "pid": 49625, "tid": 49625, "ts": 7253740911048.215,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911057.401, "dur": 4.165,
    "args": {
      "External id": 2114,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1066,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1066, "pid": 49625, "tid": 49625, "ts": 7253740911057.401,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911066.461, "dur": 3.618,
    "args": {
      "External id": 2115,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1067,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1067, "pid": 49625, "tid": 49625, "ts": 7253740911066.461,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911073.760, "dur": 3.704,
    "args": {
      "External id": 2116,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1068,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1068, "pid": 49625, "tid": 49625, "ts": 7253740911073.760,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911081.814, "dur": 3.897,
    "args": {
      "External id": 2117,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1069,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1069, "pid": 49625, "tid": 49625, "ts": 7253740911081.814,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911090.691, "dur": 3.768,
    "args": {
      "External id": 2118,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1070,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1070, "pid": 49625, "tid": 49625, "ts": 7253740911090.691,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911098.756, "dur": 3.080,
    "args": {
      "External id": 2120,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1071,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1071, "pid": 49625, "tid": 49625, "ts": 7253740911098.756,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911105.166, "dur": 3.668,
    "args": {
      "External id": 2121,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1072,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1072, "pid": 49625, "tid": 49625, "ts": 7253740911105.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911135.366, "dur": 4.074,
    "args": {
      "External id": 2123,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1073,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1073, "pid": 49625, "tid": 49625, "ts": 7253740911135.366,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911147.592, "dur": 4.168,
    "args": {
      "External id": 2126,      "cid": 134, "correlation": 1074, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1074, "pid": 49625, "tid": 49625, "ts": 7253740911147.592,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911211.943, "dur": 4.550,
    "args": {
      "External id": 2133,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1075,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1075, "pid": 49625, "tid": 49625, "ts": 7253740911211.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911222.477, "dur": 4.141,
    "args": {
      "External id": 2134,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1076,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1076, "pid": 49625, "tid": 49625, "ts": 7253740911222.477,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911231.771, "dur": 5.145,
    "args": {
      "External id": 2135,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1077,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1077, "pid": 49625, "tid": 49625, "ts": 7253740911231.771,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911240.826, "dur": 4.302,
    "args": {
      "External id": 2136,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1078,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1078, "pid": 49625, "tid": 49625, "ts": 7253740911240.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911249.908, "dur": 4.930,
    "args": {
      "External id": 2137,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1079,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1079, "pid": 49625, "tid": 49625, "ts": 7253740911249.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911260.376, "dur": 4.008,
    "args": {
      "External id": 2138,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1080,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1080, "pid": 49625, "tid": 49625, "ts": 7253740911260.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911268.774, "dur": 3.918,
    "args": {
      "External id": 2140,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1081,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1081, "pid": 49625, "tid": 49625, "ts": 7253740911268.774,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911276.164, "dur": 3.721,
    "args": {
      "External id": 2141,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1082,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1082, "pid": 49625, "tid": 49625, "ts": 7253740911276.164,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911309.589, "dur": 3.669,
    "args": {
      "External id": 2143,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1083,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1083, "pid": 49625, "tid": 49625, "ts": 7253740911309.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911322.430, "dur": 3.915,
    "args": {
      "External id": 2146,      "cid": 134, "correlation": 1084, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1084, "pid": 49625, "tid": 49625, "ts": 7253740911322.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911375.279, "dur": 3.464,
    "args": {
      "External id": 2153,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1085,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1085, "pid": 49625, "tid": 49625, "ts": 7253740911375.279,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911383.987, "dur": 4.453,
    "args": {
      "External id": 2154,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1086,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1086, "pid": 49625, "tid": 49625, "ts": 7253740911383.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911393.447, "dur": 4.322,
    "args": {
      "External id": 2155,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1087,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1087, "pid": 49625, "tid": 49625, "ts": 7253740911393.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911401.393, "dur": 3.956,
    "args": {
      "External id": 2156,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1088,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1088, "pid": 49625, "tid": 49625, "ts": 7253740911401.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911409.870, "dur": 3.114,
    "args": {
      "External id": 2157,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1089,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1089, "pid": 49625, "tid": 49625, "ts": 7253740911409.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911418.315, "dur": 4.155,
    "args": {
      "External id": 2158,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1090,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1090, "pid": 49625, "tid": 49625, "ts": 7253740911418.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911426.777, "dur": 4.050,
    "args": {
      "External id": 2160,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1091,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1091, "pid": 49625, "tid": 49625, "ts": 7253740911426.777,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911434.310, "dur": 3.883,
    "args": {
      "External id": 2161,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1092,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1092, "pid": 49625, "tid": 49625, "ts": 7253740911434.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911465.263, "dur": 4.143,
    "args": {
      "External id": 2163,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1093,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1093, "pid": 49625, "tid": 49625, "ts": 7253740911465.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911477.917, "dur": 4.619,
    "args": {
      "External id": 2166,      "cid": 134, "correlation": 1094, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1094, "pid": 49625, "tid": 49625, "ts": 7253740911477.917,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911528.607, "dur": 4.075,
    "args": {
      "External id": 2173,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1095,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1095, "pid": 49625, "tid": 49625, "ts": 7253740911528.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911537.829, "dur": 3.826,
    "args": {
      "External id": 2174,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1096,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1096, "pid": 49625, "tid": 49625, "ts": 7253740911537.829,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911546.253, "dur": 3.181,
    "args": {
      "External id": 2175,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1097,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1097, "pid": 49625, "tid": 49625, "ts": 7253740911546.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911553.130, "dur": 3.791,
    "args": {
      "External id": 2176,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1098,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1098, "pid": 49625, "tid": 49625, "ts": 7253740911553.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911561.294, "dur": 4.223,
    "args": {
      "External id": 2177,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1099,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1099, "pid": 49625, "tid": 49625, "ts": 7253740911561.294,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911570.913, "dur": 4.125,
    "args": {
      "External id": 2178,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1100,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1100, "pid": 49625, "tid": 49625, "ts": 7253740911570.913,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911579.208, "dur": 3.687,
    "args": {
      "External id": 2180,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1101,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1101, "pid": 49625, "tid": 49625, "ts": 7253740911579.208,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911586.321, "dur": 3.571,
    "args": {
      "External id": 2181,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1102,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1102, "pid": 49625, "tid": 49625, "ts": 7253740911586.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911621.948, "dur": 4.604,
    "args": {
      "External id": 2183,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1103,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1103, "pid": 49625, "tid": 49625, "ts": 7253740911621.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911635.050, "dur": 4.618,
    "args": {
      "External id": 2186,      "cid": 134, "correlation": 1104, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1104, "pid": 49625, "tid": 49625, "ts": 7253740911635.050,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911684.937, "dur": 4.749,
    "args": {
      "External id": 2193,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1105,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1105, "pid": 49625, "tid": 49625, "ts": 7253740911684.937,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911694.871, "dur": 4.825,
    "args": {
      "External id": 2194,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1106,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1106, "pid": 49625, "tid": 49625, "ts": 7253740911694.871,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911704.454, "dur": 4.293,
    "args": {
      "External id": 2195,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1107,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1107, "pid": 49625, "tid": 49625, "ts": 7253740911704.454,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911712.564, "dur": 4.010,
    "args": {
      "External id": 2196,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1108,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1108, "pid": 49625, "tid": 49625, "ts": 7253740911712.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911720.904, "dur": 4.268,
    "args": {
      "External id": 2197,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1109,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1109, "pid": 49625, "tid": 49625, "ts": 7253740911720.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911730.476, "dur": 4.170,
    "args": {
      "External id": 2198,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1110,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1110, "pid": 49625, "tid": 49625, "ts": 7253740911730.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911738.932, "dur": 4.617,
    "args": {
      "External id": 2200,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1111,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1111, "pid": 49625, "tid": 49625, "ts": 7253740911738.932,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911746.988, "dur": 4.382,
    "args": {
      "External id": 2201,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1112,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1112, "pid": 49625, "tid": 49625, "ts": 7253740911746.988,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911778.952, "dur": 5.023,
    "args": {
      "External id": 2203,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1113,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1113, "pid": 49625, "tid": 49625, "ts": 7253740911778.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911793.229, "dur": 5.806,
    "args": {
      "External id": 2206,      "cid": 134, "correlation": 1114, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1114, "pid": 49625, "tid": 49625, "ts": 7253740911793.229,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911847.137, "dur": 5.299,
    "args": {
      "External id": 2213,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1115,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1115, "pid": 49625, "tid": 49625, "ts": 7253740911847.137,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911857.786, "dur": 4.372,
    "args": {
      "External id": 2214,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1116,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1116, "pid": 49625, "tid": 49625, "ts": 7253740911857.786,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911867.009, "dur": 4.041,
    "args": {
      "External id": 2215,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1117,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1117, "pid": 49625, "tid": 49625, "ts": 7253740911867.009,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911874.640, "dur": 4.024,
    "args": {
      "External id": 2216,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1118,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1118, "pid": 49625, "tid": 49625, "ts": 7253740911874.640,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911883.080, "dur": 4.546,
    "args": {
      "External id": 2217,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1119,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1119, "pid": 49625, "tid": 49625, "ts": 7253740911883.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911893.034, "dur": 3.939,
    "args": {
      "External id": 2218,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1120,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1120, "pid": 49625, "tid": 49625, "ts": 7253740911893.034,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911901.928, "dur": 3.425,
    "args": {
      "External id": 2220,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1121,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1121, "pid": 49625, "tid": 49625, "ts": 7253740911901.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911908.712, "dur": 4.701,
    "args": {
      "External id": 2221,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1122,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1122, "pid": 49625, "tid": 49625, "ts": 7253740911908.712,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740911940.633, "dur": 5.016,
    "args": {
      "External id": 2223,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1123,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1123, "pid": 49625, "tid": 49625, "ts": 7253740911940.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740911953.874, "dur": 4.607,
    "args": {
      "External id": 2226,      "cid": 134, "correlation": 1124, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1124, "pid": 49625, "tid": 49625, "ts": 7253740911953.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912003.146, "dur": 3.366,
    "args": {
      "External id": 2233,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1125,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1125, "pid": 49625, "tid": 49625, "ts": 7253740912003.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912011.703, "dur": 4.099,
    "args": {
      "External id": 2234,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1126,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1126, "pid": 49625, "tid": 49625, "ts": 7253740912011.703,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912020.476, "dur": 4.030,
    "args": {
      "External id": 2235,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1127,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1127, "pid": 49625, "tid": 49625, "ts": 7253740912020.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912028.096, "dur": 3.986,
    "args": {
      "External id": 2236,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1128,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1128, "pid": 49625, "tid": 49625, "ts": 7253740912028.096,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912036.400, "dur": 3.651,
    "args": {
      "External id": 2237,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1129,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1129, "pid": 49625, "tid": 49625, "ts": 7253740912036.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912045.303, "dur": 3.365,
    "args": {
      "External id": 2238,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1130,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1130, "pid": 49625, "tid": 49625, "ts": 7253740912045.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912052.714, "dur": 4.135,
    "args": {
      "External id": 2240,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1131,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1131, "pid": 49625, "tid": 49625, "ts": 7253740912052.714,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912060.264, "dur": 4.107,
    "args": {
      "External id": 2241,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1132,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1132, "pid": 49625, "tid": 49625, "ts": 7253740912060.264,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912090.185, "dur": 3.938,
    "args": {
      "External id": 2243,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1133,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1133, "pid": 49625, "tid": 49625, "ts": 7253740912090.185,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912102.256, "dur": 4.623,
    "args": {
      "External id": 2246,      "cid": 134, "correlation": 1134, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1134, "pid": 49625, "tid": 49625, "ts": 7253740912102.256,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912152.403, "dur": 4.620,
    "args": {
      "External id": 2253,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1135,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1135, "pid": 49625, "tid": 49625, "ts": 7253740912152.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912162.393, "dur": 3.670,
    "args": {
      "External id": 2254,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1136,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1136, "pid": 49625, "tid": 49625, "ts": 7253740912162.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912170.684, "dur": 3.563,
    "args": {
      "External id": 2255,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1137,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1137, "pid": 49625, "tid": 49625, "ts": 7253740912170.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912177.800, "dur": 3.739,
    "args": {
      "External id": 2256,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1138,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1138, "pid": 49625, "tid": 49625, "ts": 7253740912177.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912185.836, "dur": 3.285,
    "args": {
      "External id": 2257,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1139,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1139, "pid": 49625, "tid": 49625, "ts": 7253740912185.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912194.272, "dur": 4.045,
    "args": {
      "External id": 2258,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1140,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1140, "pid": 49625, "tid": 49625, "ts": 7253740912194.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912202.465, "dur": 3.482,
    "args": {
      "External id": 2260,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1141,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1141, "pid": 49625, "tid": 49625, "ts": 7253740912202.465,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912209.364, "dur": 4.126,
    "args": {
      "External id": 2261,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1142,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1142, "pid": 49625, "tid": 49625, "ts": 7253740912209.364,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912239.213, "dur": 4.144,
    "args": {
      "External id": 2263,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1143,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1143, "pid": 49625, "tid": 49625, "ts": 7253740912239.213,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912251.310, "dur": 4.088,
    "args": {
      "External id": 2266,      "cid": 134, "correlation": 1144, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1144, "pid": 49625, "tid": 49625, "ts": 7253740912251.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912298.430, "dur": 4.635,
    "args": {
      "External id": 2273,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1145,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1145, "pid": 49625, "tid": 49625, "ts": 7253740912298.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912308.480, "dur": 3.862,
    "args": {
      "External id": 2274,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1146,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1146, "pid": 49625, "tid": 49625, "ts": 7253740912308.480,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912316.889, "dur": 3.679,
    "args": {
      "External id": 2275,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1147,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1147, "pid": 49625, "tid": 49625, "ts": 7253740912316.889,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912324.167, "dur": 3.657,
    "args": {
      "External id": 2276,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1148,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1148, "pid": 49625, "tid": 49625, "ts": 7253740912324.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912332.155, "dur": 3.504,
    "args": {
      "External id": 2277,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1149,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1149, "pid": 49625, "tid": 49625, "ts": 7253740912332.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912340.803, "dur": 4.144,
    "args": {
      "External id": 2278,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1150,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1150, "pid": 49625, "tid": 49625, "ts": 7253740912340.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912350.275, "dur": 11.736,
    "args": {
      "External id": 2280,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1151,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1151, "pid": 49625, "tid": 49625, "ts": 7253740912350.275,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740907313.651, "dur": 79.702,
    "args": {
      "External id": 1097,      "device": 2, "stream": 0,      "correlation": 559, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 559, "pid": 2, "tid": 0, "ts": 7253740907313.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907393.354, "dur": 65.470,
    "args": {
      "External id": 1098,      "device": 2, "stream": 0,      "correlation": 560, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 560, "pid": 2, "tid": 0, "ts": 7253740907393.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912370.639, "dur": 5.120,
    "args": {
      "External id": 2281,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1152,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1152, "pid": 49625, "tid": 49625, "ts": 7253740912370.639,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740907458.825, "dur": 78.379,
    "args": {
      "External id": 1100,      "device": 2, "stream": 0,      "correlation": 561, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 561, "pid": 2, "tid": 0, "ts": 7253740907458.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740907537.205, "dur": 84.833,
    "args": {
      "External id": 1101,      "device": 2, "stream": 0,      "correlation": 562, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 562, "pid": 2, "tid": 0, "ts": 7253740907537.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907629.136, "dur": 44.982,
    "args": {
      "External id": 1103,      "device": 2, "stream": 0,      "correlation": 563, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 563, "pid": 2, "tid": 0, "ts": 7253740907629.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740907674.119, "dur": 78.740,
    "args": {
      "External id": 1106,      "device": 2, "stream": 0,      "correlation": 564, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1704.5469069480957
    }
  },
  {
    "ph": "f", "id": 564, "pid": 2, "tid": 0, "ts": 7253740907674.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907760.959, "dur": 5.852,
    "args": {
      "External id": 1113,      "device": 2, "stream": 0,      "correlation": 565, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 565, "pid": 2, "tid": 0, "ts": 7253740907760.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907766.812, "dur": 6.253,
    "args": {
      "External id": 1114,      "device": 2, "stream": 0,      "correlation": 566, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 566, "pid": 2, "tid": 0, "ts": 7253740907766.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740907773.066, "dur": 85.115,
    "args": {
      "External id": 1115,      "device": 2, "stream": 0,      "correlation": 567, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 567, "pid": 2, "tid": 0, "ts": 7253740907773.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907858.182, "dur": 62.222,
    "args": {
      "External id": 1116,      "device": 2, "stream": 0,      "correlation": 568, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 568, "pid": 2, "tid": 0, "ts": 7253740907858.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740907920.405, "dur": 78.098,
    "args": {
      "External id": 1117,      "device": 2, "stream": 0,      "correlation": 569, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 569, "pid": 2, "tid": 0, "ts": 7253740907920.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740907998.504, "dur": 64.708,
    "args": {
      "External id": 1118,      "device": 2, "stream": 0,      "correlation": 570, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 570, "pid": 2, "tid": 0, "ts": 7253740907998.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912408.818, "dur": 4.902,
    "args": {
      "External id": 2283,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1153,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1153, "pid": 49625, "tid": 49625, "ts": 7253740912408.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740908063.213, "dur": 77.898,
    "args": {
      "External id": 1120,      "device": 2, "stream": 0,      "correlation": 571, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 571, "pid": 2, "tid": 0, "ts": 7253740908063.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740908141.112, "dur": 85.194,
    "args": {
      "External id": 1121,      "device": 2, "stream": 0,      "correlation": 572, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 572, "pid": 2, "tid": 0, "ts": 7253740908141.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908233.444, "dur": 45.944,
    "args": {
      "External id": 1123,      "device": 2, "stream": 0,      "correlation": 573, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 573, "pid": 2, "tid": 0, "ts": 7253740908233.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740908279.389, "dur": 77.618,
    "args": {
      "External id": 1126,      "device": 2, "stream": 0,      "correlation": 574, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1729.1865136113581
    }
  },
  {
    "ph": "f", "id": 574, "pid": 2, "tid": 0, "ts": 7253740908279.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908364.224, "dur": 5.492,
    "args": {
      "External id": 1133,      "device": 2, "stream": 0,      "correlation": 575, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 575, "pid": 2, "tid": 0, "ts": 7253740908364.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912423.631, "dur": 4.634,
    "args": {
      "External id": 2286,      "cid": 134, "correlation": 1154, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1154, "pid": 49625, "tid": 49625, "ts": 7253740912423.631,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908369.717, "dur": 5.531,
    "args": {
      "External id": 1134,      "device": 2, "stream": 0,      "correlation": 576, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 576, "pid": 2, "tid": 0, "ts": 7253740908369.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740908375.249, "dur": 85.115,
    "args": {
      "External id": 1135,      "device": 2, "stream": 0,      "correlation": 577, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 577, "pid": 2, "tid": 0, "ts": 7253740908375.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908460.365, "dur": 58.934,
    "args": {
      "External id": 1136,      "device": 2, "stream": 0,      "correlation": 578, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 578, "pid": 2, "tid": 0, "ts": 7253740908460.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740908519.300, "dur": 78.580,
    "args": {
      "External id": 1137,      "device": 2, "stream": 0,      "correlation": 579, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 579, "pid": 2, "tid": 0, "ts": 7253740908519.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908597.881, "dur": 65.710,
    "args": {
      "External id": 1138,      "device": 2, "stream": 0,      "correlation": 580, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 580, "pid": 2, "tid": 0, "ts": 7253740908597.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740908663.592, "dur": 77.296,
    "args": {
      "External id": 1140,      "device": 2, "stream": 0,      "correlation": 581, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 581, "pid": 2, "tid": 0, "ts": 7253740908663.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740908740.889, "dur": 83.391,
    "args": {
      "External id": 1141,      "device": 2, "stream": 0,      "correlation": 582, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 582, "pid": 2, "tid": 0, "ts": 7253740908740.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908832.179, "dur": 45.102,
    "args": {
      "External id": 1143,      "device": 2, "stream": 0,      "correlation": 583, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 583, "pid": 2, "tid": 0, "ts": 7253740908832.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740908877.282, "dur": 77.137,
    "args": {
      "External id": 1146,      "device": 2, "stream": 0,      "correlation": 584, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1739.9689906401513
    }
  },
  {
    "ph": "f", "id": 584, "pid": 2, "tid": 0, "ts": 7253740908877.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908961.155, "dur": 5.412,
    "args": {
      "External id": 1153,      "device": 2, "stream": 0,      "correlation": 585, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 585, "pid": 2, "tid": 0, "ts": 7253740908961.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740908966.568, "dur": 6.133,
    "args": {
      "External id": 1154,      "device": 2, "stream": 0,      "correlation": 586, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 586, "pid": 2, "tid": 0, "ts": 7253740908966.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740908972.702, "dur": 82.027,
    "args": {
      "External id": 1155,      "device": 2, "stream": 0,      "correlation": 587, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 587, "pid": 2, "tid": 0, "ts": 7253740908972.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909054.730, "dur": 60.859,
    "args": {
      "External id": 1156,      "device": 2, "stream": 0,      "correlation": 588, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 588, "pid": 2, "tid": 0, "ts": 7253740909054.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740909115.590, "dur": 77.256,
    "args": {
      "External id": 1157,      "device": 2, "stream": 0,      "correlation": 589, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 589, "pid": 2, "tid": 0, "ts": 7253740909115.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912478.648, "dur": 4.315,
    "args": {
      "External id": 2293,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1155,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1155, "pid": 49625, "tid": 49625, "ts": 7253740912478.648,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909192.847, "dur": 64.106,
    "args": {
      "External id": 1158,      "device": 2, "stream": 0,      "correlation": 590, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 590, "pid": 2, "tid": 0, "ts": 7253740909192.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740909256.954, "dur": 75.373,
    "args": {
      "External id": 1160,      "device": 2, "stream": 0,      "correlation": 591, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 591, "pid": 2, "tid": 0, "ts": 7253740909256.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912489.258, "dur": 3.637,
    "args": {
      "External id": 2294,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1156,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1156, "pid": 49625, "tid": 49625, "ts": 7253740912489.258,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740909332.328, "dur": 82.508,
    "args": {
      "External id": 1161,      "device": 2, "stream": 0,      "correlation": 592, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 592, "pid": 2, "tid": 0, "ts": 7253740909332.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909421.773, "dur": 44.662,
    "args": {
      "External id": 1163,      "device": 2, "stream": 0,      "correlation": 593, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 593, "pid": 2, "tid": 0, "ts": 7253740909421.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740909466.436, "dur": 77.456,
    "args": {
      "External id": 1166,      "device": 2, "stream": 0,      "correlation": 594, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1732.8030778367352
    }
  },
  {
    "ph": "f", "id": 594, "pid": 2, "tid": 0, "ts": 7253740909466.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909550.789, "dur": 5.291,
    "args": {
      "External id": 1173,      "device": 2, "stream": 0,      "correlation": 595, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 595, "pid": 2, "tid": 0, "ts": 7253740909550.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912497.885, "dur": 5.206,
    "args": {
      "External id": 2295,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1157,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1157, "pid": 49625, "tid": 49625, "ts": 7253740912497.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909556.081, "dur": 5.893,
    "args": {
      "External id": 1174,      "device": 2, "stream": 0,      "correlation": 596, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 596, "pid": 2, "tid": 0, "ts": 7253740909556.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740909561.975, "dur": 82.108,
    "args": {
      "External id": 1175,      "device": 2, "stream": 0,      "correlation": 597, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 597, "pid": 2, "tid": 0, "ts": 7253740909561.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912507.090, "dur": 3.586,
    "args": {
      "External id": 2296,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1158,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1158, "pid": 49625, "tid": 49625, "ts": 7253740912507.090,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909644.084, "dur": 59.495,
    "args": {
      "External id": 1176,      "device": 2, "stream": 0,      "correlation": 598, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 598, "pid": 2, "tid": 0, "ts": 7253740909644.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740909703.580, "dur": 76.655,
    "args": {
      "External id": 1177,      "device": 2, "stream": 0,      "correlation": 599, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 599, "pid": 2, "tid": 0, "ts": 7253740909703.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912515.404, "dur": 5.716,
    "args": {
      "External id": 2297,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1159,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1159, "pid": 49625, "tid": 49625, "ts": 7253740912515.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740909780.236, "dur": 66.071,
    "args": {
      "External id": 1178,      "device": 2, "stream": 0,      "correlation": 600, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 600, "pid": 2, "tid": 0, "ts": 7253740909780.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740909846.308, "dur": 75.653,
    "args": {
      "External id": 1180,      "device": 2, "stream": 0,      "correlation": 601, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 601, "pid": 2, "tid": 0, "ts": 7253740909846.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912527.033, "dur": 4.349,
    "args": {
      "External id": 2298,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1160,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1160, "pid": 49625, "tid": 49625, "ts": 7253740912527.033,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740909921.962, "dur": 81.827,
    "args": {
      "External id": 1181,      "device": 2, "stream": 0,      "correlation": 602, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 602, "pid": 2, "tid": 0, "ts": 7253740909921.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910010.806, "dur": 44.301,
    "args": {
      "External id": 1183,      "device": 2, "stream": 0,      "correlation": 603, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 603, "pid": 2, "tid": 0, "ts": 7253740910010.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740910055.108, "dur": 77.898,
    "args": {
      "External id": 1186,      "device": 2, "stream": 0,      "correlation": 604, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1722.9711292827892
    }
  },
  {
    "ph": "f", "id": 604, "pid": 2, "tid": 0, "ts": 7253740910055.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910140.383, "dur": 5.332,
    "args": {
      "External id": 1193,      "device": 2, "stream": 0,      "correlation": 605, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 605, "pid": 2, "tid": 0, "ts": 7253740910140.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912536.163, "dur": 4.008,
    "args": {
      "External id": 2300,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1161,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1161, "pid": 49625, "tid": 49625, "ts": 7253740912536.163,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910145.716, "dur": 5.411,
    "args": {
      "External id": 1194,      "device": 2, "stream": 0,      "correlation": 606, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 606, "pid": 2, "tid": 0, "ts": 7253740910145.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740910151.128, "dur": 82.308,
    "args": {
      "External id": 1195,      "device": 2, "stream": 0,      "correlation": 607, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 607, "pid": 2, "tid": 0, "ts": 7253740910151.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912543.809, "dur": 3.935,
    "args": {
      "External id": 2301,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1162,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1162, "pid": 49625, "tid": 49625, "ts": 7253740912543.809,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910233.437, "dur": 61.380,
    "args": {
      "External id": 1196,      "device": 2, "stream": 0,      "correlation": 608, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 608, "pid": 2, "tid": 0, "ts": 7253740910233.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740910294.818, "dur": 77.377,
    "args": {
      "External id": 1197,      "device": 2, "stream": 0,      "correlation": 609, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 609, "pid": 2, "tid": 0, "ts": 7253740910294.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910372.196, "dur": 67.634,
    "args": {
      "External id": 1198,      "device": 2, "stream": 0,      "correlation": 610, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 610, "pid": 2, "tid": 0, "ts": 7253740910372.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740910439.831, "dur": 76.094,
    "args": {
      "External id": 1200,      "device": 2, "stream": 0,      "correlation": 611, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 611, "pid": 2, "tid": 0, "ts": 7253740910439.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740910515.926, "dur": 81.787,
    "args": {
      "External id": 1201,      "device": 2, "stream": 0,      "correlation": 612, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 612, "pid": 2, "tid": 0, "ts": 7253740910515.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910605.452, "dur": 44.661,
    "args": {
      "External id": 1203,      "device": 2, "stream": 0,      "correlation": 613, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 613, "pid": 2, "tid": 0, "ts": 7253740910605.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740910650.114, "dur": 79.502,
    "args": {
      "External id": 1206,      "device": 2, "stream": 0,      "correlation": 614, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1688.2096021533778
    }
  },
  {
    "ph": "f", "id": 614, "pid": 2, "tid": 0, "ts": 7253740910650.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910736.432, "dur": 5.372,
    "args": {
      "External id": 1213,      "device": 2, "stream": 0,      "correlation": 615, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 615, "pid": 2, "tid": 0, "ts": 7253740910736.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910741.805, "dur": 5.612,
    "args": {
      "External id": 1214,      "device": 2, "stream": 0,      "correlation": 616, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 616, "pid": 2, "tid": 0, "ts": 7253740910741.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912575.808, "dur": 5.223,
    "args": {
      "External id": 2303,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1163,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1163, "pid": 49625, "tid": 49625, "ts": 7253740912575.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740910747.418, "dur": 81.466,
    "args": {
      "External id": 1215,      "device": 2, "stream": 0,      "correlation": 617, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 617, "pid": 2, "tid": 0, "ts": 7253740910747.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910828.885, "dur": 62.583,
    "args": {
      "External id": 1216,      "device": 2, "stream": 0,      "correlation": 618, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 618, "pid": 2, "tid": 0, "ts": 7253740910828.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740910891.469, "dur": 75.933,
    "args": {
      "External id": 1217,      "device": 2, "stream": 0,      "correlation": 619, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 619, "pid": 2, "tid": 0, "ts": 7253740910891.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912589.533, "dur": 4.798,
    "args": {
      "External id": 2306,      "cid": 134, "correlation": 1164, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1164, "pid": 49625, "tid": 49625, "ts": 7253740912589.533,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740910967.403, "dur": 64.427,
    "args": {
      "External id": 1218,      "device": 2, "stream": 0,      "correlation": 620, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 620, "pid": 2, "tid": 0, "ts": 7253740910967.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740911031.831, "dur": 74.771,
    "args": {
      "External id": 1220,      "device": 2, "stream": 0,      "correlation": 621, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 621, "pid": 2, "tid": 0, "ts": 7253740911031.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740911106.603, "dur": 79.982,
    "args": {
      "External id": 1221,      "device": 2, "stream": 0,      "correlation": 622, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 622, "pid": 2, "tid": 0, "ts": 7253740911106.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911193.602, "dur": 45.704,
    "args": {
      "External id": 1223,      "device": 2, "stream": 0,      "correlation": 623, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 623, "pid": 2, "tid": 0, "ts": 7253740911193.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740911239.307, "dur": 78.379,
    "args": {
      "External id": 1226,      "device": 2, "stream": 0,      "correlation": 624, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1712.3976524623629
    }
  },
  {
    "ph": "f", "id": 624, "pid": 2, "tid": 0, "ts": 7253740911239.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911324.583, "dur": 5.211,
    "args": {
      "External id": 1233,      "device": 2, "stream": 0,      "correlation": 625, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 625, "pid": 2, "tid": 0, "ts": 7253740911324.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911329.795, "dur": 5.291,
    "args": {
      "External id": 1234,      "device": 2, "stream": 0,      "correlation": 626, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 626, "pid": 2, "tid": 0, "ts": 7253740911329.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740911335.087, "dur": 79.702,
    "args": {
      "External id": 1235,      "device": 2, "stream": 0,      "correlation": 627, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 627, "pid": 2, "tid": 0, "ts": 7253740911335.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911414.790, "dur": 62.503,
    "args": {
      "External id": 1236,      "device": 2, "stream": 0,      "correlation": 628, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 628, "pid": 2, "tid": 0, "ts": 7253740911414.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740911477.294, "dur": 76.094,
    "args": {
      "External id": 1237,      "device": 2, "stream": 0,      "correlation": 629, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 629, "pid": 2, "tid": 0, "ts": 7253740911477.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911553.389, "dur": 63.545,
    "args": {
      "External id": 1238,      "device": 2, "stream": 0,      "correlation": 630, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 630, "pid": 2, "tid": 0, "ts": 7253740911553.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740911616.935, "dur": 73.047,
    "args": {
      "External id": 1240,      "device": 2, "stream": 0,      "correlation": 631, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 631, "pid": 2, "tid": 0, "ts": 7253740911616.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740911689.983, "dur": 79.261,
    "args": {
      "External id": 1241,      "device": 2, "stream": 0,      "correlation": 632, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 632, "pid": 2, "tid": 0, "ts": 7253740911689.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911776.461, "dur": 45.183,
    "args": {
      "External id": 1243,      "device": 2, "stream": 0,      "correlation": 633, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 633, "pid": 2, "tid": 0, "ts": 7253740911776.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740911821.645, "dur": 75.372,
    "args": {
      "External id": 1246,      "device": 2, "stream": 0,      "correlation": 634, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1780.7136242421027
    }
  },
  {
    "ph": "f", "id": 634, "pid": 2, "tid": 0, "ts": 7253740911821.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912642.999, "dur": 3.698,
    "args": {
      "External id": 2313,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1165,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1165, "pid": 49625, "tid": 49625, "ts": 7253740912642.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911904.435, "dur": 5.171,
    "args": {
      "External id": 1253,      "device": 2, "stream": 0,      "correlation": 635, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 635, "pid": 2, "tid": 0, "ts": 7253740911904.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911909.607, "dur": 5.812,
    "args": {
      "External id": 1254,      "device": 2, "stream": 0,      "correlation": 636, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 636, "pid": 2, "tid": 0, "ts": 7253740911909.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912652.285, "dur": 4.302,
    "args": {
      "External id": 2314,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1166,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1166, "pid": 49625, "tid": 49625, "ts": 7253740912652.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740911915.420, "dur": 78.219,
    "args": {
      "External id": 1255,      "device": 2, "stream": 0,      "correlation": 637, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 637, "pid": 2, "tid": 0, "ts": 7253740911915.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912662.099, "dur": 3.711,
    "args": {
      "External id": 2315,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1167,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1167, "pid": 49625, "tid": 49625, "ts": 7253740912662.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912669.770, "dur": 3.685,
    "args": {
      "External id": 2316,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1168,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1168, "pid": 49625, "tid": 49625, "ts": 7253740912669.770,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912677.982, "dur": 4.008,
    "args": {
      "External id": 2317,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1169,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1169, "pid": 49625, "tid": 49625, "ts": 7253740912677.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912687.209, "dur": 4.390,
    "args": {
      "External id": 2318,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1170,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1170, "pid": 49625, "tid": 49625, "ts": 7253740912687.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912695.874, "dur": 4.176,
    "args": {
      "External id": 2320,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1171,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1171, "pid": 49625, "tid": 49625, "ts": 7253740912695.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912703.473, "dur": 3.563,
    "args": {
      "External id": 2321,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1172,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1172, "pid": 49625, "tid": 49625, "ts": 7253740912703.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912735.116, "dur": 5.368,
    "args": {
      "External id": 2323,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1173,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1173, "pid": 49625, "tid": 49625, "ts": 7253740912735.116,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912748.544, "dur": 4.508,
    "args": {
      "External id": 2326,      "cid": 134, "correlation": 1174, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1174, "pid": 49625, "tid": 49625, "ts": 7253740912748.544,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912799.073, "dur": 4.228,
    "args": {
      "External id": 2333,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1175,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1175, "pid": 49625, "tid": 49625, "ts": 7253740912799.073,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912808.881, "dur": 3.533,
    "args": {
      "External id": 2334,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1176,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1176, "pid": 49625, "tid": 49625, "ts": 7253740912808.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912817.096, "dur": 3.883,
    "args": {
      "External id": 2335,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1177,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1177, "pid": 49625, "tid": 49625, "ts": 7253740912817.096,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912824.511, "dur": 3.738,
    "args": {
      "External id": 2336,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1178,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1178, "pid": 49625, "tid": 49625, "ts": 7253740912824.511,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912832.556, "dur": 3.346,
    "args": {
      "External id": 2337,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1179,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1179, "pid": 49625, "tid": 49625, "ts": 7253740912832.556,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912841.029, "dur": 4.085,
    "args": {
      "External id": 2338,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1180,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1180, "pid": 49625, "tid": 49625, "ts": 7253740912841.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912849.369, "dur": 3.971,
    "args": {
      "External id": 2340,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1181,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1181, "pid": 49625, "tid": 49625, "ts": 7253740912849.369,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912856.692, "dur": 3.737,
    "args": {
      "External id": 2341,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1182,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1182, "pid": 49625, "tid": 49625, "ts": 7253740912856.692,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912888.027, "dur": 4.079,
    "args": {
      "External id": 2343,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1183,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1183, "pid": 49625, "tid": 49625, "ts": 7253740912888.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740912900.390, "dur": 4.244,
    "args": {
      "External id": 2346,      "cid": 134, "correlation": 1184, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1184, "pid": 49625, "tid": 49625, "ts": 7253740912900.390,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912955.856, "dur": 4.602,
    "args": {
      "External id": 2353,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1185,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1185, "pid": 49625, "tid": 49625, "ts": 7253740912955.856,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912966.131, "dur": 4.943,
    "args": {
      "External id": 2354,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1186,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1186, "pid": 49625, "tid": 49625, "ts": 7253740912966.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912975.853, "dur": 4.171,
    "args": {
      "External id": 2355,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1187,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1187, "pid": 49625, "tid": 49625, "ts": 7253740912975.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912983.668, "dur": 4.298,
    "args": {
      "External id": 2356,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1188,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1188, "pid": 49625, "tid": 49625, "ts": 7253740912983.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740912992.360, "dur": 4.026,
    "args": {
      "External id": 2357,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1189,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1189, "pid": 49625, "tid": 49625, "ts": 7253740912992.360,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913001.554, "dur": 3.945,
    "args": {
      "External id": 2358,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1190,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1190, "pid": 49625, "tid": 49625, "ts": 7253740913001.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913009.746, "dur": 3.607,
    "args": {
      "External id": 2360,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1191,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1191, "pid": 49625, "tid": 49625, "ts": 7253740913009.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913016.726, "dur": 3.896,
    "args": {
      "External id": 2361,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1192,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1192, "pid": 49625, "tid": 49625, "ts": 7253740913016.726,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913047.801, "dur": 3.420,
    "args": {
      "External id": 2363,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1193,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1193, "pid": 49625, "tid": 49625, "ts": 7253740913047.801,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913060.416, "dur": 4.496,
    "args": {
      "External id": 2366,      "cid": 134, "correlation": 1194, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1194, "pid": 49625, "tid": 49625, "ts": 7253740913060.416,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913110.193, "dur": 3.625,
    "args": {
      "External id": 2373,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1195,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1195, "pid": 49625, "tid": 49625, "ts": 7253740913110.193,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913119.119, "dur": 3.941,
    "args": {
      "External id": 2374,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1196,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1196, "pid": 49625, "tid": 49625, "ts": 7253740913119.119,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913127.786, "dur": 3.578,
    "args": {
      "External id": 2375,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1197,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1197, "pid": 49625, "tid": 49625, "ts": 7253740913127.786,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913135.000, "dur": 3.510,
    "args": {
      "External id": 2376,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1198,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1198, "pid": 49625, "tid": 49625, "ts": 7253740913135.000,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913142.881, "dur": 3.758,
    "args": {
      "External id": 2377,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1199,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1199, "pid": 49625, "tid": 49625, "ts": 7253740913142.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913151.861, "dur": 4.391,
    "args": {
      "External id": 2378,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1200,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1200, "pid": 49625, "tid": 49625, "ts": 7253740913151.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913160.604, "dur": 3.566,
    "args": {
      "External id": 2380,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1201,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1201, "pid": 49625, "tid": 49625, "ts": 7253740913160.604,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913167.525, "dur": 3.686,
    "args": {
      "External id": 2381,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1202,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1202, "pid": 49625, "tid": 49625, "ts": 7253740913167.525,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913197.383, "dur": 4.668,
    "args": {
      "External id": 2383,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1203,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1203, "pid": 49625, "tid": 49625, "ts": 7253740913197.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913210.342, "dur": 3.765,
    "args": {
      "External id": 2386,      "cid": 134, "correlation": 1204, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1204, "pid": 49625, "tid": 49625, "ts": 7253740913210.342,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913259.009, "dur": 3.839,
    "args": {
      "External id": 2393,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1205,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1205, "pid": 49625, "tid": 49625, "ts": 7253740913259.009,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913268.110, "dur": 3.890,
    "args": {
      "External id": 2394,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1206,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1206, "pid": 49625, "tid": 49625, "ts": 7253740913268.110,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913276.672, "dur": 3.112,
    "args": {
      "External id": 2395,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1207,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1207, "pid": 49625, "tid": 49625, "ts": 7253740913276.672,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913283.414, "dur": 3.759,
    "args": {
      "External id": 2396,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1208,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1208, "pid": 49625, "tid": 49625, "ts": 7253740913283.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913291.505, "dur": 4.750,
    "args": {
      "External id": 2397,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1209,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1209, "pid": 49625, "tid": 49625, "ts": 7253740913291.505,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913301.744, "dur": 3.768,
    "args": {
      "External id": 2398,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1210,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1210, "pid": 49625, "tid": 49625, "ts": 7253740913301.744,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913309.658, "dur": 4.579,
    "args": {
      "External id": 2400,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1211,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1211, "pid": 49625, "tid": 49625, "ts": 7253740913309.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913317.560, "dur": 3.974,
    "args": {
      "External id": 2401,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1212,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1212, "pid": 49625, "tid": 49625, "ts": 7253740913317.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913346.920, "dur": 4.665,
    "args": {
      "External id": 2403,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1213,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1213, "pid": 49625, "tid": 49625, "ts": 7253740913346.920,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913359.919, "dur": 4.862,
    "args": {
      "External id": 2406,      "cid": 134, "correlation": 1214, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1214, "pid": 49625, "tid": 49625, "ts": 7253740913359.919,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913408.551, "dur": 4.380,
    "args": {
      "External id": 2413,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1215,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1215, "pid": 49625, "tid": 49625, "ts": 7253740913408.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913418.168, "dur": 3.572,
    "args": {
      "External id": 2414,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1216,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1216, "pid": 49625, "tid": 49625, "ts": 7253740913418.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913426.316, "dur": 4.181,
    "args": {
      "External id": 2415,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1217,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1217, "pid": 49625, "tid": 49625, "ts": 7253740913426.316,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913434.152, "dur": 3.786,
    "args": {
      "External id": 2416,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1218,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1218, "pid": 49625, "tid": 49625, "ts": 7253740913434.152,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913442.304, "dur": 3.843,
    "args": {
      "External id": 2417,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1219,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1219, "pid": 49625, "tid": 49625, "ts": 7253740913442.304,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913451.209, "dur": 3.767,
    "args": {
      "External id": 2418,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1220,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1220, "pid": 49625, "tid": 49625, "ts": 7253740913451.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913459.222, "dur": 3.812,
    "args": {
      "External id": 2420,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1221,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1221, "pid": 49625, "tid": 49625, "ts": 7253740913459.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913466.561, "dur": 4.347,
    "args": {
      "External id": 2421,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1222,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1222, "pid": 49625, "tid": 49625, "ts": 7253740913466.561,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913497.451, "dur": 3.910,
    "args": {
      "External id": 2423,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1223,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1223, "pid": 49625, "tid": 49625, "ts": 7253740913497.451,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913513.051, "dur": 10.400,
    "args": {
      "External id": 2426,      "cid": 134, "correlation": 1224, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1224, "pid": 49625, "tid": 49625, "ts": 7253740913513.051,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913582.894, "dur": 18.929,
    "args": {
      "External id": 2433,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1225,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1225, "pid": 49625, "tid": 49625, "ts": 7253740913582.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913609.208, "dur": 4.790,
    "args": {
      "External id": 2434,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1226,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1226, "pid": 49625, "tid": 49625, "ts": 7253740913609.208,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913619.352, "dur": 4.475,
    "args": {
      "External id": 2435,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1227,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1227, "pid": 49625, "tid": 49625, "ts": 7253740913619.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913627.785, "dur": 3.961,
    "args": {
      "External id": 2436,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1228,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1228, "pid": 49625, "tid": 49625, "ts": 7253740913627.785,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913637.374, "dur": 3.888,
    "args": {
      "External id": 2437,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1229,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1229, "pid": 49625, "tid": 49625, "ts": 7253740913637.374,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913647.189, "dur": 5.032,
    "args": {
      "External id": 2438,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1230,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1230, "pid": 49625, "tid": 49625, "ts": 7253740913647.189,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913656.977, "dur": 3.008,
    "args": {
      "External id": 2440,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1231,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1231, "pid": 49625, "tid": 49625, "ts": 7253740913656.977,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913663.438, "dur": 3.956,
    "args": {
      "External id": 2441,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1232,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1232, "pid": 49625, "tid": 49625, "ts": 7253740913663.438,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913697.411, "dur": 4.644,
    "args": {
      "External id": 2443,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1233,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1233, "pid": 49625, "tid": 49625, "ts": 7253740913697.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913711.663, "dur": 4.988,
    "args": {
      "External id": 2446,      "cid": 134, "correlation": 1234, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1234, "pid": 49625, "tid": 49625, "ts": 7253740913711.663,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913762.586, "dur": 4.393,
    "args": {
      "External id": 2453,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1235,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1235, "pid": 49625, "tid": 49625, "ts": 7253740913762.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913772.405, "dur": 3.763,
    "args": {
      "External id": 2454,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1236,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1236, "pid": 49625, "tid": 49625, "ts": 7253740913772.405,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913781.156, "dur": 3.704,
    "args": {
      "External id": 2455,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1237,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1237, "pid": 49625, "tid": 49625, "ts": 7253740913781.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913788.489, "dur": 3.173,
    "args": {
      "External id": 2456,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1238,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1238, "pid": 49625, "tid": 49625, "ts": 7253740913788.489,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913796.066, "dur": 4.860,
    "args": {
      "External id": 2457,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1239,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1239, "pid": 49625, "tid": 49625, "ts": 7253740913796.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913806.303, "dur": 3.880,
    "args": {
      "External id": 2458,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1240,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1240, "pid": 49625, "tid": 49625, "ts": 7253740913806.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913814.466, "dur": 3.671,
    "args": {
      "External id": 2460,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1241,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1241, "pid": 49625, "tid": 49625, "ts": 7253740913814.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913821.643, "dur": 3.587,
    "args": {
      "External id": 2461,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1242,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1242, "pid": 49625, "tid": 49625, "ts": 7253740913821.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913852.686, "dur": 3.996,
    "args": {
      "External id": 2463,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1243,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1243, "pid": 49625, "tid": 49625, "ts": 7253740913852.686,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740913865.367, "dur": 3.506,
    "args": {
      "External id": 2466,      "cid": 134, "correlation": 1244, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1244, "pid": 49625, "tid": 49625, "ts": 7253740913865.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913914.398, "dur": 3.899,
    "args": {
      "External id": 2473,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1245,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1245, "pid": 49625, "tid": 49625, "ts": 7253740913914.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913923.781, "dur": 3.298,
    "args": {
      "External id": 2474,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1246,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1246, "pid": 49625, "tid": 49625, "ts": 7253740913923.781,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913931.626, "dur": 3.560,
    "args": {
      "External id": 2475,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1247,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1247, "pid": 49625, "tid": 49625, "ts": 7253740913931.626,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913938.928, "dur": 3.807,
    "args": {
      "External id": 2476,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1248,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1248, "pid": 49625, "tid": 49625, "ts": 7253740913938.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913947.011, "dur": 3.632,
    "args": {
      "External id": 2477,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1249,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1249, "pid": 49625, "tid": 49625, "ts": 7253740913947.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913955.876, "dur": 3.623,
    "args": {
      "External id": 2478,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1250,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1250, "pid": 49625, "tid": 49625, "ts": 7253740913955.876,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913963.695, "dur": 4.028,
    "args": {
      "External id": 2480,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1251,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1251, "pid": 49625, "tid": 49625, "ts": 7253740913963.695,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740913971.095, "dur": 3.571,
    "args": {
      "External id": 2481,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1252,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1252, "pid": 49625, "tid": 49625, "ts": 7253740913971.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914001.311, "dur": 3.689,
    "args": {
      "External id": 2483,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1253,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1253, "pid": 49625, "tid": 49625, "ts": 7253740914001.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914013.458, "dur": 5.529,
    "args": {
      "External id": 2486,      "cid": 134, "correlation": 1254, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1254, "pid": 49625, "tid": 49625, "ts": 7253740914013.458,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914062.793, "dur": 4.013,
    "args": {
      "External id": 2493,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1255,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1255, "pid": 49625, "tid": 49625, "ts": 7253740914062.793,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914108.120, "dur": 4.508,
    "args": {
      "External id": 2494,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1256,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1256, "pid": 49625, "tid": 49625, "ts": 7253740914108.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914119.870, "dur": 3.698,
    "args": {
      "External id": 2495,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1257,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1257, "pid": 49625, "tid": 49625, "ts": 7253740914119.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914128.190, "dur": 4.783,
    "args": {
      "External id": 2496,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1258,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1258, "pid": 49625, "tid": 49625, "ts": 7253740914128.190,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914138.162, "dur": 5.196,
    "args": {
      "External id": 2497,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1259,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1259, "pid": 49625, "tid": 49625, "ts": 7253740914138.162,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914149.677, "dur": 5.053,
    "args": {
      "External id": 2498,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1260,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1260, "pid": 49625, "tid": 49625, "ts": 7253740914149.677,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914159.473, "dur": 3.912,
    "args": {
      "External id": 2500,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1261,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1261, "pid": 49625, "tid": 49625, "ts": 7253740914159.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914166.874, "dur": 4.126,
    "args": {
      "External id": 2501,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1262,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1262, "pid": 49625, "tid": 49625, "ts": 7253740914166.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914204.432, "dur": 4.283,
    "args": {
      "External id": 2503,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1263,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1263, "pid": 49625, "tid": 49625, "ts": 7253740914204.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914218.800, "dur": 4.875,
    "args": {
      "External id": 2506,      "cid": 134, "correlation": 1264, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1264, "pid": 49625, "tid": 49625, "ts": 7253740914218.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914274.190, "dur": 6.368,
    "args": {
      "External id": 2513,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1265,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1265, "pid": 49625, "tid": 49625, "ts": 7253740914274.190,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914286.714, "dur": 3.573,
    "args": {
      "External id": 2514,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1266,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1266, "pid": 49625, "tid": 49625, "ts": 7253740914286.714,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914295.015, "dur": 3.811,
    "args": {
      "External id": 2515,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1267,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1267, "pid": 49625, "tid": 49625, "ts": 7253740914295.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914302.615, "dur": 3.830,
    "args": {
      "External id": 2516,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1268,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1268, "pid": 49625, "tid": 49625, "ts": 7253740914302.615,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914310.915, "dur": 3.360,
    "args": {
      "External id": 2517,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1269,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1269, "pid": 49625, "tid": 49625, "ts": 7253740914310.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914319.648, "dur": 4.154,
    "args": {
      "External id": 2518,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1270,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1270, "pid": 49625, "tid": 49625, "ts": 7253740914319.648,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914328.340, "dur": 4.452,
    "args": {
      "External id": 2520,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1271,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1271, "pid": 49625, "tid": 49625, "ts": 7253740914328.340,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914336.302, "dur": 4.061,
    "args": {
      "External id": 2521,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1272,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1272, "pid": 49625, "tid": 49625, "ts": 7253740914336.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914369.184, "dur": 5.272,
    "args": {
      "External id": 2523,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1273,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1273, "pid": 49625, "tid": 49625, "ts": 7253740914369.184,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914383.127, "dur": 4.370,
    "args": {
      "External id": 2526,      "cid": 134, "correlation": 1274, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1274, "pid": 49625, "tid": 49625, "ts": 7253740914383.127,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914434.197, "dur": 4.152,
    "args": {
      "External id": 2533,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1275,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1275, "pid": 49625, "tid": 49625, "ts": 7253740914434.197,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914443.873, "dur": 4.373,
    "args": {
      "External id": 2534,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1276,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1276, "pid": 49625, "tid": 49625, "ts": 7253740914443.873,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914452.910, "dur": 3.423,
    "args": {
      "External id": 2535,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1277,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1277, "pid": 49625, "tid": 49625, "ts": 7253740914452.910,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914459.972, "dur": 3.894,
    "args": {
      "External id": 2536,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1278,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1278, "pid": 49625, "tid": 49625, "ts": 7253740914459.972,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914468.284, "dur": 4.409,
    "args": {
      "External id": 2537,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1279,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1279, "pid": 49625, "tid": 49625, "ts": 7253740914468.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914478.015, "dur": 3.900,
    "args": {
      "External id": 2538,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1280,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1280, "pid": 49625, "tid": 49625, "ts": 7253740914478.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914486.182, "dur": 3.522,
    "args": {
      "External id": 2540,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1281,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1281, "pid": 49625, "tid": 49625, "ts": 7253740914486.182,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914493.180, "dur": 3.861,
    "args": {
      "External id": 2541,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1282,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1282, "pid": 49625, "tid": 49625, "ts": 7253740914493.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914523.670, "dur": 4.393,
    "args": {
      "External id": 2543,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1283,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1283, "pid": 49625, "tid": 49625, "ts": 7253740914523.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914536.655, "dur": 4.967,
    "args": {
      "External id": 2546,      "cid": 134, "correlation": 1284, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1284, "pid": 49625, "tid": 49625, "ts": 7253740914536.655,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914586.102, "dur": 3.548,
    "args": {
      "External id": 2553,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1285,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1285, "pid": 49625, "tid": 49625, "ts": 7253740914586.102,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914595.122, "dur": 3.806,
    "args": {
      "External id": 2554,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1286,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1286, "pid": 49625, "tid": 49625, "ts": 7253740914595.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914603.491, "dur": 4.763,
    "args": {
      "External id": 2555,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1287,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1287, "pid": 49625, "tid": 49625, "ts": 7253740914603.491,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914612.064, "dur": 3.781,
    "args": {
      "External id": 2556,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1288,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1288, "pid": 49625, "tid": 49625, "ts": 7253740914612.064,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914620.183, "dur": 3.813,
    "args": {
      "External id": 2557,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1289,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1289, "pid": 49625, "tid": 49625, "ts": 7253740914620.183,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914629.199, "dur": 3.644,
    "args": {
      "External id": 2558,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1290,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1290, "pid": 49625, "tid": 49625, "ts": 7253740914629.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914636.961, "dur": 4.110,
    "args": {
      "External id": 2560,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1291,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1291, "pid": 49625, "tid": 49625, "ts": 7253740914636.961,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914652.860, "dur": 3.663,
    "args": {
      "External id": 2561,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1292,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1292, "pid": 49625, "tid": 49625, "ts": 7253740914652.860,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914682.560, "dur": 10.757,
    "args": {
      "External id": 2563,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1293,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1293, "pid": 49625, "tid": 49625, "ts": 7253740914682.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914707.022, "dur": 6.344,
    "args": {
      "External id": 2566,      "cid": 134, "correlation": 1294, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1294, "pid": 49625, "tid": 49625, "ts": 7253740914707.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914768.754, "dur": 4.835,
    "args": {
      "External id": 2573,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1295,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1295, "pid": 49625, "tid": 49625, "ts": 7253740914768.754,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914780.141, "dur": 4.237,
    "args": {
      "External id": 2574,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1296,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1296, "pid": 49625, "tid": 49625, "ts": 7253740914780.141,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914789.402, "dur": 3.920,
    "args": {
      "External id": 2575,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1297,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1297, "pid": 49625, "tid": 49625, "ts": 7253740914789.402,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914797.118, "dur": 3.418,
    "args": {
      "External id": 2576,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1298,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1298, "pid": 49625, "tid": 49625, "ts": 7253740914797.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914805.155, "dur": 4.539,
    "args": {
      "External id": 2577,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1299,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1299, "pid": 49625, "tid": 49625, "ts": 7253740914805.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914816.625, "dur": 5.284,
    "args": {
      "External id": 2578,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1300,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1300, "pid": 49625, "tid": 49625, "ts": 7253740914816.625,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914826.575, "dur": 3.135,
    "args": {
      "External id": 2580,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1301,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1301, "pid": 49625, "tid": 49625, "ts": 7253740914826.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914833.118, "dur": 8.496,
    "args": {
      "External id": 2581,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1302,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1302, "pid": 49625, "tid": 49625, "ts": 7253740914833.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914876.891, "dur": 4.488,
    "args": {
      "External id": 2583,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1303,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1303, "pid": 49625, "tid": 49625, "ts": 7253740914876.891,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740914891.491, "dur": 4.665,
    "args": {
      "External id": 2586,      "cid": 134, "correlation": 1304, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1304, "pid": 49625, "tid": 49625, "ts": 7253740914891.491,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914945.924, "dur": 5.127,
    "args": {
      "External id": 2593,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1305,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1305, "pid": 49625, "tid": 49625, "ts": 7253740914945.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914957.477, "dur": 4.376,
    "args": {
      "External id": 2594,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1306,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1306, "pid": 49625, "tid": 49625, "ts": 7253740914957.477,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914966.931, "dur": 3.671,
    "args": {
      "External id": 2595,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1307,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1307, "pid": 49625, "tid": 49625, "ts": 7253740914966.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914974.411, "dur": 4.537,
    "args": {
      "External id": 2596,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1308,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1308, "pid": 49625, "tid": 49625, "ts": 7253740914974.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914983.840, "dur": 4.269,
    "args": {
      "External id": 2597,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1309,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1309, "pid": 49625, "tid": 49625, "ts": 7253740914983.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740914993.572, "dur": 4.208,
    "args": {
      "External id": 2598,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1310,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1310, "pid": 49625, "tid": 49625, "ts": 7253740914993.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915002.345, "dur": 3.826,
    "args": {
      "External id": 2600,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1311,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1311, "pid": 49625, "tid": 49625, "ts": 7253740915002.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915009.637, "dur": 3.701,
    "args": {
      "External id": 2601,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1312,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1312, "pid": 49625, "tid": 49625, "ts": 7253740915009.637,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915042.367, "dur": 4.039,
    "args": {
      "External id": 2603,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1313,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1313, "pid": 49625, "tid": 49625, "ts": 7253740915042.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915054.904, "dur": 4.986,
    "args": {
      "External id": 2606,      "cid": 134, "correlation": 1314, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1314, "pid": 49625, "tid": 49625, "ts": 7253740915054.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915105.476, "dur": 3.485,
    "args": {
      "External id": 2613,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1315,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1315, "pid": 49625, "tid": 49625, "ts": 7253740915105.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915114.273, "dur": 3.773,
    "args": {
      "External id": 2614,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1316,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1316, "pid": 49625, "tid": 49625, "ts": 7253740915114.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915122.968, "dur": 3.986,
    "args": {
      "External id": 2615,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1317,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1317, "pid": 49625, "tid": 49625, "ts": 7253740915122.968,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915130.600, "dur": 3.737,
    "args": {
      "External id": 2616,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1318,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1318, "pid": 49625, "tid": 49625, "ts": 7253740915130.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915138.897, "dur": 4.646,
    "args": {
      "External id": 2617,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1319,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1319, "pid": 49625, "tid": 49625, "ts": 7253740915138.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915148.771, "dur": 3.958,
    "args": {
      "External id": 2618,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1320,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1320, "pid": 49625, "tid": 49625, "ts": 7253740915148.771,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915156.968, "dur": 4.749,
    "args": {
      "External id": 2620,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1321,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1321, "pid": 49625, "tid": 49625, "ts": 7253740915156.968,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915165.100, "dur": 4.009,
    "args": {
      "External id": 2621,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1322,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1322, "pid": 49625, "tid": 49625, "ts": 7253740915165.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915196.134, "dur": 4.588,
    "args": {
      "External id": 2623,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1323,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1323, "pid": 49625, "tid": 49625, "ts": 7253740915196.134,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915209.176, "dur": 4.408,
    "args": {
      "External id": 2626,      "cid": 134, "correlation": 1324, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1324, "pid": 49625, "tid": 49625, "ts": 7253740915209.176,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915259.488, "dur": 3.632,
    "args": {
      "External id": 2633,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1325,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1325, "pid": 49625, "tid": 49625, "ts": 7253740915259.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915268.358, "dur": 6.103,
    "args": {
      "External id": 2634,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1326,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1326, "pid": 49625, "tid": 49625, "ts": 7253740915268.358,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915284.303, "dur": 9.419,
    "args": {
      "External id": 2635,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1327,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1327, "pid": 49625, "tid": 49625, "ts": 7253740915284.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915301.877, "dur": 5.063,
    "args": {
      "External id": 2636,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1328,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1328, "pid": 49625, "tid": 49625, "ts": 7253740915301.877,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915313.564, "dur": 4.748,
    "args": {
      "External id": 2637,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1329,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1329, "pid": 49625, "tid": 49625, "ts": 7253740915313.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915324.606, "dur": 4.714,
    "args": {
      "External id": 2638,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1330,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1330, "pid": 49625, "tid": 49625, "ts": 7253740915324.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915333.817, "dur": 4.464,
    "args": {
      "External id": 2640,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1331,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1331, "pid": 49625, "tid": 49625, "ts": 7253740915333.817,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915342.440, "dur": 4.086,
    "args": {
      "External id": 2641,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1332,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1332, "pid": 49625, "tid": 49625, "ts": 7253740915342.440,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915377.780, "dur": 4.255,
    "args": {
      "External id": 2643,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1333,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1333, "pid": 49625, "tid": 49625, "ts": 7253740915377.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915391.542, "dur": 4.163,
    "args": {
      "External id": 2646,      "cid": 134, "correlation": 1334, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1334, "pid": 49625, "tid": 49625, "ts": 7253740915391.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915445.923, "dur": 5.476,
    "args": {
      "External id": 2653,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1335,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1335, "pid": 49625, "tid": 49625, "ts": 7253740915445.923,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915457.515, "dur": 3.537,
    "args": {
      "External id": 2654,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1336,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1336, "pid": 49625, "tid": 49625, "ts": 7253740915457.515,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915465.847, "dur": 3.736,
    "args": {
      "External id": 2655,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1337,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1337, "pid": 49625, "tid": 49625, "ts": 7253740915465.847,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915473.352, "dur": 3.755,
    "args": {
      "External id": 2656,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1338,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1338, "pid": 49625, "tid": 49625, "ts": 7253740915473.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915481.572, "dur": 3.533,
    "args": {
      "External id": 2657,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1339,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1339, "pid": 49625, "tid": 49625, "ts": 7253740915481.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915490.440, "dur": 3.396,
    "args": {
      "External id": 2658,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1340,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1340, "pid": 49625, "tid": 49625, "ts": 7253740915490.440,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915498.052, "dur": 3.836,
    "args": {
      "External id": 2660,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1341,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1341, "pid": 49625, "tid": 49625, "ts": 7253740915498.052,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915505.257, "dur": 3.918,
    "args": {
      "External id": 2661,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1342,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1342, "pid": 49625, "tid": 49625, "ts": 7253740915505.257,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915537.740, "dur": 4.628,
    "args": {
      "External id": 2663,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1343,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1343, "pid": 49625, "tid": 49625, "ts": 7253740915537.740,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915550.926, "dur": 4.421,
    "args": {
      "External id": 2666,      "cid": 134, "correlation": 1344, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1344, "pid": 49625, "tid": 49625, "ts": 7253740915550.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915601.949, "dur": 4.534,
    "args": {
      "External id": 2673,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1345,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1345, "pid": 49625, "tid": 49625, "ts": 7253740915601.949,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915611.959, "dur": 3.685,
    "args": {
      "External id": 2674,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1346,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1346, "pid": 49625, "tid": 49625, "ts": 7253740915611.959,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915620.256, "dur": 3.549,
    "args": {
      "External id": 2675,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1347,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1347, "pid": 49625, "tid": 49625, "ts": 7253740915620.256,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915627.421, "dur": 3.358,
    "args": {
      "External id": 2676,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1348,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1348, "pid": 49625, "tid": 49625, "ts": 7253740915627.421,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915635.236, "dur": 3.756,
    "args": {
      "External id": 2677,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1349,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1349, "pid": 49625, "tid": 49625, "ts": 7253740915635.236,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915644.315, "dur": 3.986,
    "args": {
      "External id": 2678,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1350,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1350, "pid": 49625, "tid": 49625, "ts": 7253740915644.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915652.488, "dur": 3.465,
    "args": {
      "External id": 2680,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1351,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1351, "pid": 49625, "tid": 49625, "ts": 7253740915652.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915659.389, "dur": 3.435,
    "args": {
      "External id": 2681,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1352,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1352, "pid": 49625, "tid": 49625, "ts": 7253740915659.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915690.863, "dur": 3.367,
    "args": {
      "External id": 2683,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1353,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1353, "pid": 49625, "tid": 49625, "ts": 7253740915690.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915702.500, "dur": 4.181,
    "args": {
      "External id": 2686,      "cid": 134, "correlation": 1354, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1354, "pid": 49625, "tid": 49625, "ts": 7253740915702.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915753.039, "dur": 4.232,
    "args": {
      "External id": 2693,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1355,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1355, "pid": 49625, "tid": 49625, "ts": 7253740915753.039,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915762.621, "dur": 3.894,
    "args": {
      "External id": 2694,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1356,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1356, "pid": 49625, "tid": 49625, "ts": 7253740915762.621,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915771.122, "dur": 3.309,
    "args": {
      "External id": 2695,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1357,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1357, "pid": 49625, "tid": 49625, "ts": 7253740915771.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915778.033, "dur": 4.270,
    "args": {
      "External id": 2696,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1358,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1358, "pid": 49625, "tid": 49625, "ts": 7253740915778.033,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915786.896, "dur": 4.278,
    "args": {
      "External id": 2697,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1359,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1359, "pid": 49625, "tid": 49625, "ts": 7253740915786.896,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915796.381, "dur": 4.110,
    "args": {
      "External id": 2698,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1360,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1360, "pid": 49625, "tid": 49625, "ts": 7253740915796.381,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915804.662, "dur": 3.758,
    "args": {
      "External id": 2700,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1361,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1361, "pid": 49625, "tid": 49625, "ts": 7253740915804.662,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915811.725, "dur": 3.493,
    "args": {
      "External id": 2701,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1362,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1362, "pid": 49625, "tid": 49625, "ts": 7253740915811.725,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915842.310, "dur": 4.371,
    "args": {
      "External id": 2703,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1363,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1363, "pid": 49625, "tid": 49625, "ts": 7253740915842.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740915855.037, "dur": 5.020,
    "args": {
      "External id": 2706,      "cid": 134, "correlation": 1364, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1364, "pid": 49625, "tid": 49625, "ts": 7253740915855.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915923.496, "dur": 6.216,
    "args": {
      "External id": 2713,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1365,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1365, "pid": 49625, "tid": 49625, "ts": 7253740915923.496,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915936.830, "dur": 4.510,
    "args": {
      "External id": 2714,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1366,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1366, "pid": 49625, "tid": 49625, "ts": 7253740915936.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915946.685, "dur": 4.569,
    "args": {
      "External id": 2715,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1367,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1367, "pid": 49625, "tid": 49625, "ts": 7253740915946.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915955.031, "dur": 4.678,
    "args": {
      "External id": 2716,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1368,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1368, "pid": 49625, "tid": 49625, "ts": 7253740915955.031,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915964.489, "dur": 4.328,
    "args": {
      "External id": 2717,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1369,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1369, "pid": 49625, "tid": 49625, "ts": 7253740915964.489,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915974.460, "dur": 4.453,
    "args": {
      "External id": 2718,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1370,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1370, "pid": 49625, "tid": 49625, "ts": 7253740915974.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915983.504, "dur": 4.250,
    "args": {
      "External id": 2720,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1371,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1371, "pid": 49625, "tid": 49625, "ts": 7253740915983.504,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740915996.932, "dur": 4.501,
    "args": {
      "External id": 2721,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1372,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1372, "pid": 49625, "tid": 49625, "ts": 7253740915996.932,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916037.181, "dur": 4.378,
    "args": {
      "External id": 2723,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1373,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1373, "pid": 49625, "tid": 49625, "ts": 7253740916037.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916051.517, "dur": 5.190,
    "args": {
      "External id": 2726,      "cid": 134, "correlation": 1374, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1374, "pid": 49625, "tid": 49625, "ts": 7253740916051.517,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916107.019, "dur": 3.880,
    "args": {
      "External id": 2733,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1375,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1375, "pid": 49625, "tid": 49625, "ts": 7253740916107.019,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916117.227, "dur": 4.420,
    "args": {
      "External id": 2734,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1376,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1376, "pid": 49625, "tid": 49625, "ts": 7253740916117.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916126.588, "dur": 4.245,
    "args": {
      "External id": 2735,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1377,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1377, "pid": 49625, "tid": 49625, "ts": 7253740916126.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916134.645, "dur": 4.441,
    "args": {
      "External id": 2736,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1378,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1378, "pid": 49625, "tid": 49625, "ts": 7253740916134.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916143.821, "dur": 3.435,
    "args": {
      "External id": 2737,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1379,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1379, "pid": 49625, "tid": 49625, "ts": 7253740916143.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916152.740, "dur": 4.057,
    "args": {
      "External id": 2738,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1380,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1380, "pid": 49625, "tid": 49625, "ts": 7253740916152.740,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916161.211, "dur": 4.455,
    "args": {
      "External id": 2740,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1381,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1381, "pid": 49625, "tid": 49625, "ts": 7253740916161.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916169.156, "dur": 3.693,
    "args": {
      "External id": 2741,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1382,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1382, "pid": 49625, "tid": 49625, "ts": 7253740916169.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916201.404, "dur": 4.409,
    "args": {
      "External id": 2743,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1383,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1383, "pid": 49625, "tid": 49625, "ts": 7253740916201.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916214.314, "dur": 5.581,
    "args": {
      "External id": 2746,      "cid": 134, "correlation": 1384, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1384, "pid": 49625, "tid": 49625, "ts": 7253740916214.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916266.612, "dur": 4.654,
    "args": {
      "External id": 2753,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1385,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1385, "pid": 49625, "tid": 49625, "ts": 7253740916266.612,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916276.964, "dur": 3.760,
    "args": {
      "External id": 2754,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1386,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1386, "pid": 49625, "tid": 49625, "ts": 7253740916276.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916285.488, "dur": 3.382,
    "args": {
      "External id": 2755,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1387,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1387, "pid": 49625, "tid": 49625, "ts": 7253740916285.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916292.569, "dur": 3.712,
    "args": {
      "External id": 2756,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1388,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1388, "pid": 49625, "tid": 49625, "ts": 7253740916292.569,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916300.864, "dur": 3.957,
    "args": {
      "External id": 2757,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1389,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1389, "pid": 49625, "tid": 49625, "ts": 7253740916300.864,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916310.057, "dur": 3.741,
    "args": {
      "External id": 2758,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1390,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1390, "pid": 49625, "tid": 49625, "ts": 7253740916310.057,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916317.911, "dur": 3.336,
    "args": {
      "External id": 2760,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1391,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1391, "pid": 49625, "tid": 49625, "ts": 7253740916317.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916324.629, "dur": 3.672,
    "args": {
      "External id": 2761,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1392,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1392, "pid": 49625, "tid": 49625, "ts": 7253740916324.629,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916355.883, "dur": 4.263,
    "args": {
      "External id": 2763,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1393,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1393, "pid": 49625, "tid": 49625, "ts": 7253740916355.883,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916368.663, "dur": 4.155,
    "args": {
      "External id": 2766,      "cid": 134, "correlation": 1394, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1394, "pid": 49625, "tid": 49625, "ts": 7253740916368.663,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916419.564, "dur": 3.970,
    "args": {
      "External id": 2773,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1395,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1395, "pid": 49625, "tid": 49625, "ts": 7253740916419.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916428.937, "dur": 3.543,
    "args": {
      "External id": 2774,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1396,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1396, "pid": 49625, "tid": 49625, "ts": 7253740916428.937,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916437.061, "dur": 3.687,
    "args": {
      "External id": 2775,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1397,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1397, "pid": 49625, "tid": 49625, "ts": 7253740916437.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916444.393, "dur": 3.157,
    "args": {
      "External id": 2776,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1398,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1398, "pid": 49625, "tid": 49625, "ts": 7253740916444.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916452.066, "dur": 6.338,
    "args": {
      "External id": 2777,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1399,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1399, "pid": 49625, "tid": 49625, "ts": 7253740916452.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916468.873, "dur": 8.930,
    "args": {
      "External id": 2778,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1400,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1400, "pid": 49625, "tid": 49625, "ts": 7253740916468.873,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916487.645, "dur": 4.545,
    "args": {
      "External id": 2780,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1401,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1401, "pid": 49625, "tid": 49625, "ts": 7253740916487.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916496.891, "dur": 4.204,
    "args": {
      "External id": 2781,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1402,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1402, "pid": 49625, "tid": 49625, "ts": 7253740916496.891,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916533.831, "dur": 4.315,
    "args": {
      "External id": 2783,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1403,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1403, "pid": 49625, "tid": 49625, "ts": 7253740916533.831,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916547.786, "dur": 4.456,
    "args": {
      "External id": 2786,      "cid": 134, "correlation": 1404, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1404, "pid": 49625, "tid": 49625, "ts": 7253740916547.786,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916602.789, "dur": 4.193,
    "args": {
      "External id": 2793,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1405,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1405, "pid": 49625, "tid": 49625, "ts": 7253740916602.789,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916612.878, "dur": 3.526,
    "args": {
      "External id": 2794,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1406,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1406, "pid": 49625, "tid": 49625, "ts": 7253740916612.878,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916621.117, "dur": 4.565,
    "args": {
      "External id": 2795,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1407,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1407, "pid": 49625, "tid": 49625, "ts": 7253740916621.117,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916629.602, "dur": 3.380,
    "args": {
      "External id": 2796,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1408,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1408, "pid": 49625, "tid": 49625, "ts": 7253740916629.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916637.546, "dur": 4.136,
    "args": {
      "External id": 2797,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1409,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1409, "pid": 49625, "tid": 49625, "ts": 7253740916637.546,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916647.281, "dur": 3.619,
    "args": {
      "External id": 2798,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1410,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1410, "pid": 49625, "tid": 49625, "ts": 7253740916647.281,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916655.093, "dur": 4.051,
    "args": {
      "External id": 2800,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1411,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1411, "pid": 49625, "tid": 49625, "ts": 7253740916655.093,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916662.602, "dur": 3.779,
    "args": {
      "External id": 2801,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1412,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1412, "pid": 49625, "tid": 49625, "ts": 7253740916662.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916694.511, "dur": 5.642,
    "args": {
      "External id": 2803,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1413,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1413, "pid": 49625, "tid": 49625, "ts": 7253740916694.511,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916708.936, "dur": 4.563,
    "args": {
      "External id": 2806,      "cid": 134, "correlation": 1414, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1414, "pid": 49625, "tid": 49625, "ts": 7253740916708.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916760.610, "dur": 3.535,
    "args": {
      "External id": 2813,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1415,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1415, "pid": 49625, "tid": 49625, "ts": 7253740916760.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916769.572, "dur": 4.756,
    "args": {
      "External id": 2814,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1416,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1416, "pid": 49625, "tid": 49625, "ts": 7253740916769.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916779.020, "dur": 3.831,
    "args": {
      "External id": 2815,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1417,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1417, "pid": 49625, "tid": 49625, "ts": 7253740916779.020,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916786.627, "dur": 3.881,
    "args": {
      "External id": 2816,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1418,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1418, "pid": 49625, "tid": 49625, "ts": 7253740916786.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916794.869, "dur": 3.884,
    "args": {
      "External id": 2817,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1419,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1419, "pid": 49625, "tid": 49625, "ts": 7253740916794.869,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916803.936, "dur": 3.811,
    "args": {
      "External id": 2818,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1420,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1420, "pid": 49625, "tid": 49625, "ts": 7253740916803.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916811.908, "dur": 3.311,
    "args": {
      "External id": 2820,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1421,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1421, "pid": 49625, "tid": 49625, "ts": 7253740916811.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916818.513, "dur": 4.033,
    "args": {
      "External id": 2821,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1422,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1422, "pid": 49625, "tid": 49625, "ts": 7253740916818.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916849.649, "dur": 4.351,
    "args": {
      "External id": 2823,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1423,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1423, "pid": 49625, "tid": 49625, "ts": 7253740916849.649,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740916862.392, "dur": 4.480,
    "args": {
      "External id": 2826,      "cid": 134, "correlation": 1424, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1424, "pid": 49625, "tid": 49625, "ts": 7253740916862.392,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916914.109, "dur": 5.351,
    "args": {
      "External id": 2833,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1425,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1425, "pid": 49625, "tid": 49625, "ts": 7253740916914.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916925.078, "dur": 4.151,
    "args": {
      "External id": 2834,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1426,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1426, "pid": 49625, "tid": 49625, "ts": 7253740916925.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916933.796, "dur": 3.921,
    "args": {
      "External id": 2835,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1427,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1427, "pid": 49625, "tid": 49625, "ts": 7253740916933.796,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916941.470, "dur": 3.995,
    "args": {
      "External id": 2836,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1428,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1428, "pid": 49625, "tid": 49625, "ts": 7253740916941.470,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916949.983, "dur": 3.479,
    "args": {
      "External id": 2837,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1429,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1429, "pid": 49625, "tid": 49625, "ts": 7253740916949.983,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916958.741, "dur": 3.975,
    "args": {
      "External id": 2838,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1430,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1430, "pid": 49625, "tid": 49625, "ts": 7253740916958.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916967.098, "dur": 4.066,
    "args": {
      "External id": 2840,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1431,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1431, "pid": 49625, "tid": 49625, "ts": 7253740916967.098,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740916974.542, "dur": 3.759,
    "args": {
      "External id": 2841,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1432,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1432, "pid": 49625, "tid": 49625, "ts": 7253740916974.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917005.151, "dur": 4.508,
    "args": {
      "External id": 2843,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1433,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1433, "pid": 49625, "tid": 49625, "ts": 7253740917005.151,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740917017.890, "dur": 4.358,
    "args": {
      "External id": 2846,      "cid": 134, "correlation": 1434, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1434, "pid": 49625, "tid": 49625, "ts": 7253740917017.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740911993.640, "dur": 61.340,
    "args": {
      "External id": 1256,      "device": 2, "stream": 0,      "correlation": 638, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 638, "pid": 2, "tid": 0, "ts": 7253740911993.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740912054.981, "dur": 74.891,
    "args": {
      "External id": 1257,      "device": 2, "stream": 0,      "correlation": 639, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 639, "pid": 2, "tid": 0, "ts": 7253740912054.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917083.251, "dur": 4.852,
    "args": {
      "External id": 2853,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1435,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1435, "pid": 49625, "tid": 49625, "ts": 7253740917083.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912129.873, "dur": 62.101,
    "args": {
      "External id": 1258,      "device": 2, "stream": 0,      "correlation": 640, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 640, "pid": 2, "tid": 0, "ts": 7253740912129.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740912191.975, "dur": 73.248,
    "args": {
      "External id": 1260,      "device": 2, "stream": 0,      "correlation": 641, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 641, "pid": 2, "tid": 0, "ts": 7253740912191.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917095.722, "dur": 6.235,
    "args": {
      "External id": 2854,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1436,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1436, "pid": 49625, "tid": 49625, "ts": 7253740917095.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740912265.224, "dur": 78.860,
    "args": {
      "External id": 1261,      "device": 2, "stream": 0,      "correlation": 642, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 642, "pid": 2, "tid": 0, "ts": 7253740912265.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912350.981, "dur": 44.902,
    "args": {
      "External id": 1263,      "device": 2, "stream": 0,      "correlation": 643, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 643, "pid": 2, "tid": 0, "ts": 7253740912350.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740912395.884, "dur": 77.376,
    "args": {
      "External id": 1266,      "device": 2, "stream": 0,      "correlation": 644, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1734.594621140649
    }
  },
  {
    "ph": "f", "id": 644, "pid": 2, "tid": 0, "ts": 7253740912395.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912480.037, "dur": 5.131,
    "args": {
      "External id": 1273,      "device": 2, "stream": 0,      "correlation": 645, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 645, "pid": 2, "tid": 0, "ts": 7253740912480.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917107.596, "dur": 5.018,
    "args": {
      "External id": 2855,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1437,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1437, "pid": 49625, "tid": 49625, "ts": 7253740917107.596,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912485.169, "dur": 5.211,
    "args": {
      "External id": 1274,      "device": 2, "stream": 0,      "correlation": 646, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 646, "pid": 2, "tid": 0, "ts": 7253740912485.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740912490.381, "dur": 78.138,
    "args": {
      "External id": 1275,      "device": 2, "stream": 0,      "correlation": 647, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 647, "pid": 2, "tid": 0, "ts": 7253740912490.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917116.689, "dur": 4.567,
    "args": {
      "External id": 2856,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1438,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1438, "pid": 49625, "tid": 49625, "ts": 7253740917116.689,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912568.520, "dur": 62.864,
    "args": {
      "External id": 1276,      "device": 2, "stream": 0,      "correlation": 648, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 648, "pid": 2, "tid": 0, "ts": 7253740912568.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740912631.385, "dur": 74.851,
    "args": {
      "External id": 1277,      "device": 2, "stream": 0,      "correlation": 649, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 649, "pid": 2, "tid": 0, "ts": 7253740912631.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912706.237, "dur": 65.669,
    "args": {
      "External id": 1278,      "device": 2, "stream": 0,      "correlation": 650, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 650, "pid": 2, "tid": 0, "ts": 7253740912706.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917126.235, "dur": 18.424,
    "args": {
      "External id": 2857,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1439,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1439, "pid": 49625, "tid": 49625, "ts": 7253740917126.235,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740912771.907, "dur": 72.927,
    "args": {
      "External id": 1280,      "device": 2, "stream": 0,      "correlation": 651, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 651, "pid": 2, "tid": 0, "ts": 7253740912771.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740912844.835, "dur": 77.577,
    "args": {
      "External id": 1281,      "device": 2, "stream": 0,      "correlation": 652, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 652, "pid": 2, "tid": 0, "ts": 7253740912844.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740912929.269, "dur": 45.824,
    "args": {
      "External id": 1283,      "device": 2, "stream": 0,      "correlation": 653, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 653, "pid": 2, "tid": 0, "ts": 7253740912929.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740912975.094, "dur": 78.179,
    "args": {
      "External id": 1286,      "device": 2, "stream": 0,      "correlation": 654, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1716.7783064722435
    }
  },
  {
    "ph": "f", "id": 654, "pid": 2, "tid": 0, "ts": 7253740912975.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917151.131, "dur": 4.356,
    "args": {
      "External id": 2858,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1440,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1440, "pid": 49625, "tid": 49625, "ts": 7253740917151.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913059.889, "dur": 5.091,
    "args": {
      "External id": 1293,      "device": 2, "stream": 0,      "correlation": 655, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 655, "pid": 2, "tid": 0, "ts": 7253740913059.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913064.981, "dur": 5.571,
    "args": {
      "External id": 1294,      "device": 2, "stream": 0,      "correlation": 656, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 656, "pid": 2, "tid": 0, "ts": 7253740913064.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917160.250, "dur": 4.250,
    "args": {
      "External id": 2860,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1441,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1441, "pid": 49625, "tid": 49625, "ts": 7253740917160.250,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740913070.553, "dur": 78.219,
    "args": {
      "External id": 1295,      "device": 2, "stream": 0,      "correlation": 657, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 657, "pid": 2, "tid": 0, "ts": 7253740913070.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913148.773, "dur": 62.062,
    "args": {
      "External id": 1296,      "device": 2, "stream": 0,      "correlation": 658, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 658, "pid": 2, "tid": 0, "ts": 7253740913148.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917168.384, "dur": 3.701,
    "args": {
      "External id": 2861,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1442,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1442, "pid": 49625, "tid": 49625, "ts": 7253740917168.384,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740913210.836, "dur": 77.015,
    "args": {
      "External id": 1297,      "device": 2, "stream": 0,      "correlation": 659, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 659, "pid": 2, "tid": 0, "ts": 7253740913210.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913287.852, "dur": 62.864,
    "args": {
      "External id": 1298,      "device": 2, "stream": 0,      "correlation": 660, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 660, "pid": 2, "tid": 0, "ts": 7253740913287.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740913350.717, "dur": 73.167,
    "args": {
      "External id": 1300,      "device": 2, "stream": 0,      "correlation": 661, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 661, "pid": 2, "tid": 0, "ts": 7253740913350.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740913423.885, "dur": 77.697,
    "args": {
      "External id": 1301,      "device": 2, "stream": 0,      "correlation": 662, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 662, "pid": 2, "tid": 0, "ts": 7253740913423.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913508.439, "dur": 45.303,
    "args": {
      "External id": 1303,      "device": 2, "stream": 0,      "correlation": 663, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 663, "pid": 2, "tid": 0, "ts": 7253740913508.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740913553.743, "dur": 77.858,
    "args": {
      "External id": 1306,      "device": 2, "stream": 0,      "correlation": 664, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1723.856304345034
    }
  },
  {
    "ph": "f", "id": 664, "pid": 2, "tid": 0, "ts": 7253740913553.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913639.019, "dur": 5.251,
    "args": {
      "External id": 1313,      "device": 2, "stream": 0,      "correlation": 665, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 665, "pid": 2, "tid": 0, "ts": 7253740913639.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913644.271, "dur": 5.492,
    "args": {
      "External id": 1314,      "device": 2, "stream": 0,      "correlation": 666, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 666, "pid": 2, "tid": 0, "ts": 7253740913644.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740913649.764, "dur": 78.178,
    "args": {
      "External id": 1315,      "device": 2, "stream": 0,      "correlation": 667, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 667, "pid": 2, "tid": 0, "ts": 7253740913649.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913727.943, "dur": 65.309,
    "args": {
      "External id": 1316,      "device": 2, "stream": 0,      "correlation": 668, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 668, "pid": 2, "tid": 0, "ts": 7253740913727.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917204.928, "dur": 6.004,
    "args": {
      "External id": 2863,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1443,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1443, "pid": 49625, "tid": 49625, "ts": 7253740917204.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740913793.253, "dur": 74.971,
    "args": {
      "External id": 1317,      "device": 2, "stream": 0,      "correlation": 669, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 669, "pid": 2, "tid": 0, "ts": 7253740913793.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740913868.225, "dur": 65.911,
    "args": {
      "External id": 1318,      "device": 2, "stream": 0,      "correlation": 670, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 670, "pid": 2, "tid": 0, "ts": 7253740913868.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740913934.137, "dur": 74.811,
    "args": {
      "External id": 1320,      "device": 2, "stream": 0,      "correlation": 671, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 671, "pid": 2, "tid": 0, "ts": 7253740913934.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740917220.997, "dur": 4.481,
    "args": {
      "External id": 2866,      "cid": 134, "correlation": 1444, "src": "0x7f6d4f800000", "dst": "0x7f6d57a00000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1444, "pid": 49625, "tid": 49625, "ts": 7253740917220.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740914008.949, "dur": 80.784,
    "args": {
      "External id": 1321,      "device": 2, "stream": 0,      "correlation": 672, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 672, "pid": 2, "tid": 0, "ts": 7253740914008.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914096.991, "dur": 45.904,
    "args": {
      "External id": 1323,      "device": 2, "stream": 0,      "correlation": 673, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 673, "pid": 2, "tid": 0, "ts": 7253740914096.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740914142.896, "dur": 77.617,
    "args": {
      "External id": 1326,      "device": 2, "stream": 0,      "correlation": 674, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1729.2087917751037
    }
  },
  {
    "ph": "f", "id": 674, "pid": 2, "tid": 0, "ts": 7253740914142.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914228.012, "dur": 5.371,
    "args": {
      "External id": 1333,      "device": 2, "stream": 0,      "correlation": 675, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 675, "pid": 2, "tid": 0, "ts": 7253740914228.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914233.384, "dur": 5.812,
    "args": {
      "External id": 1334,      "device": 2, "stream": 0,      "correlation": 676, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 676, "pid": 2, "tid": 0, "ts": 7253740914233.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740914239.197, "dur": 81.507,
    "args": {
      "External id": 1335,      "device": 2, "stream": 0,      "correlation": 677, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 677, "pid": 2, "tid": 0, "ts": 7253740914239.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914320.705, "dur": 63.504,
    "args": {
      "External id": 1336,      "device": 2, "stream": 0,      "correlation": 678, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 678, "pid": 2, "tid": 0, "ts": 7253740914320.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740914384.210, "dur": 76.615,
    "args": {
      "External id": 1337,      "device": 2, "stream": 0,      "correlation": 679, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 679, "pid": 2, "tid": 0, "ts": 7253740914384.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914460.826, "dur": 62.744,
    "args": {
      "External id": 1338,      "device": 2, "stream": 0,      "correlation": 680, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 680, "pid": 2, "tid": 0, "ts": 7253740914460.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740914523.571, "dur": 75.291,
    "args": {
      "External id": 1340,      "device": 2, "stream": 0,      "correlation": 681, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 681, "pid": 2, "tid": 0, "ts": 7253740914523.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740914598.863, "dur": 80.745,
    "args": {
      "External id": 1341,      "device": 2, "stream": 0,      "correlation": 682, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 682, "pid": 2, "tid": 0, "ts": 7253740914598.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914687.146, "dur": 44.582,
    "args": {
      "External id": 1343,      "device": 2, "stream": 0,      "correlation": 683, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 683, "pid": 2, "tid": 0, "ts": 7253740914687.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740914731.729, "dur": 77.777,
    "args": {
      "External id": 1346,      "device": 2, "stream": 0,      "correlation": 684, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1725.6515724240787
    }
  },
  {
    "ph": "f", "id": 684, "pid": 2, "tid": 0, "ts": 7253740914731.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914816.203, "dur": 5.371,
    "args": {
      "External id": 1353,      "device": 2, "stream": 0,      "correlation": 685, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 685, "pid": 2, "tid": 0, "ts": 7253740914816.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914821.575, "dur": 5.812,
    "args": {
      "External id": 1354,      "device": 2, "stream": 0,      "correlation": 686, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 686, "pid": 2, "tid": 0, "ts": 7253740914821.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917275.471, "dur": 3.480,
    "args": {
      "External id": 2873,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1445,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1445, "pid": 49625, "tid": 49625, "ts": 7253740917275.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740914827.388, "dur": 81.547,
    "args": {
      "External id": 1355,      "device": 2, "stream": 0,      "correlation": 687, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 687, "pid": 2, "tid": 0, "ts": 7253740914827.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740914908.936, "dur": 63.023,
    "args": {
      "External id": 1356,      "device": 2, "stream": 0,      "correlation": 688, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 688, "pid": 2, "tid": 0, "ts": 7253740914908.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917284.867, "dur": 4.722,
    "args": {
      "External id": 2874,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1446,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1446, "pid": 49625, "tid": 49625, "ts": 7253740917284.867,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740914971.960, "dur": 76.615,
    "args": {
      "External id": 1357,      "device": 2, "stream": 0,      "correlation": 689, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 689, "pid": 2, "tid": 0, "ts": 7253740914971.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917294.813, "dur": 3.715,
    "args": {
      "External id": 2875,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1447,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1447, "pid": 49625, "tid": 49625, "ts": 7253740917294.813,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915048.576, "dur": 63.585,
    "args": {
      "External id": 1358,      "device": 2, "stream": 0,      "correlation": 690, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 690, "pid": 2, "tid": 0, "ts": 7253740915048.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740915112.162, "dur": 75.453,
    "args": {
      "External id": 1360,      "device": 2, "stream": 0,      "correlation": 691, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 691, "pid": 2, "tid": 0, "ts": 7253740915112.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917302.447, "dur": 3.741,
    "args": {
      "External id": 2876,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1448,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1448, "pid": 49625, "tid": 49625, "ts": 7253740917302.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740915187.616, "dur": 82.107,
    "args": {
      "External id": 1361,      "device": 2, "stream": 0,      "correlation": 692, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 692, "pid": 2, "tid": 0, "ts": 7253740915187.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917310.657, "dur": 3.916,
    "args": {
      "External id": 2877,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1449,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1449, "pid": 49625, "tid": 49625, "ts": 7253740917310.657,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915276.580, "dur": 44.822,
    "args": {
      "External id": 1363,      "device": 2, "stream": 0,      "correlation": 693, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 693, "pid": 2, "tid": 0, "ts": 7253740915276.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740915321.403, "dur": 78.018,
    "args": {
      "External id": 1366,      "device": 2, "stream": 0,      "correlation": 694, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1720.321050000641
    }
  },
  {
    "ph": "f", "id": 694, "pid": 2, "tid": 0, "ts": 7253740915321.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915406.238, "dur": 5.411,
    "args": {
      "External id": 1373,      "device": 2, "stream": 0,      "correlation": 695, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 695, "pid": 2, "tid": 0, "ts": 7253740915406.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917320.265, "dur": 3.676,
    "args": {
      "External id": 2878,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1450,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1450, "pid": 49625, "tid": 49625, "ts": 7253740917320.265,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915411.650, "dur": 5.692,
    "args": {
      "External id": 1374,      "device": 2, "stream": 0,      "correlation": 696, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 696, "pid": 2, "tid": 0, "ts": 7253740915411.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740915417.343, "dur": 82.669,
    "args": {
      "External id": 1375,      "device": 2, "stream": 0,      "correlation": 697, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 697, "pid": 2, "tid": 0, "ts": 7253740915417.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917328.521, "dur": 5.093,
    "args": {
      "External id": 2880,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1451,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1451, "pid": 49625, "tid": 49625, "ts": 7253740917328.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915500.013, "dur": 60.939,
    "args": {
      "External id": 1376,      "device": 2, "stream": 0,      "correlation": 698, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 698, "pid": 2, "tid": 0, "ts": 7253740915500.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740915560.953, "dur": 77.377,
    "args": {
      "External id": 1377,      "device": 2, "stream": 0,      "correlation": 699, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 699, "pid": 2, "tid": 0, "ts": 7253740915560.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917337.426, "dur": 4.002,
    "args": {
      "External id": 2881,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1452,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1452, "pid": 49625, "tid": 49625, "ts": 7253740917337.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915638.331, "dur": 62.583,
    "args": {
      "External id": 1378,      "device": 2, "stream": 0,      "correlation": 700, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 700, "pid": 2, "tid": 0, "ts": 7253740915638.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740915700.915, "dur": 75.171,
    "args": {
      "External id": 1380,      "device": 2, "stream": 0,      "correlation": 701, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 701, "pid": 2, "tid": 0, "ts": 7253740915700.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740915776.087, "dur": 82.709,
    "args": {
      "External id": 1381,      "device": 2, "stream": 0,      "correlation": 702, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 702, "pid": 2, "tid": 0, "ts": 7253740915776.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915865.693, "dur": 45.784,
    "args": {
      "External id": 1383,      "device": 2, "stream": 0,      "correlation": 703, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 703, "pid": 2, "tid": 0, "ts": 7253740915865.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740915911.478, "dur": 78.219,
    "args": {
      "External id": 1386,      "device": 2, "stream": 0,      "correlation": 704, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1715.9003835336232
    }
  },
  {
    "ph": "f", "id": 704, "pid": 2, "tid": 0, "ts": 7253740915911.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740915996.794, "dur": 5.331,
    "args": {
      "External id": 1393,      "device": 2, "stream": 0,      "correlation": 705, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 705, "pid": 2, "tid": 0, "ts": 7253740915996.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916002.126, "dur": 5.853,
    "args": {
      "External id": 1394,      "device": 2, "stream": 0,      "correlation": 706, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 706, "pid": 2, "tid": 0, "ts": 7253740916002.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917370.312, "dur": 4.483,
    "args": {
      "External id": 2883,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "cid": 107, "correlation": 1453,      "grid": [65536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1453, "pid": 49625, "tid": 49625, "ts": 7253740917370.312,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740916007.980, "dur": 82.188,
    "args": {
      "External id": 1395,      "device": 2, "stream": 0,      "correlation": 707, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 707, "pid": 2, "tid": 0, "ts": 7253740916007.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916090.169, "dur": 61.821,
    "args": {
      "External id": 1396,      "device": 2, "stream": 0,      "correlation": 708, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 708, "pid": 2, "tid": 0, "ts": 7253740916090.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 49625, "tid": 49625,
    "ts": 7253740917383.685, "dur": 4.318,
    "args": {
      "External id": 2886,      "cid": 134, "correlation": 1454, "src": "0x7f6d57a00000", "dst": "0x7f6d4f800000", "bytes": "134217728", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 1454, "pid": 49625, "tid": 49625, "ts": 7253740917383.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740916151.991, "dur": 76.855,
    "args": {
      "External id": 1397,      "device": 2, "stream": 0,      "correlation": 709, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 709, "pid": 2, "tid": 0, "ts": 7253740916151.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916228.847, "dur": 66.793,
    "args": {
      "External id": 1398,      "device": 2, "stream": 0,      "correlation": 710, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 710, "pid": 2, "tid": 0, "ts": 7253740916228.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740916295.641, "dur": 74.690,
    "args": {
      "External id": 1400,      "device": 2, "stream": 0,      "correlation": 711, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 711, "pid": 2, "tid": 0, "ts": 7253740916295.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740916370.332, "dur": 82.388,
    "args": {
      "External id": 1401,      "device": 2, "stream": 0,      "correlation": 712, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 712, "pid": 2, "tid": 0, "ts": 7253740916370.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916460.259, "dur": 45.583,
    "args": {
      "External id": 1403,      "device": 2, "stream": 0,      "correlation": 713, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 713, "pid": 2, "tid": 0, "ts": 7253740916460.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740916505.843, "dur": 78.219,
    "args": {
      "External id": 1406,      "device": 2, "stream": 0,      "correlation": 714, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1715.9003835336232
    }
  },
  {
    "ph": "f", "id": 714, "pid": 2, "tid": 0, "ts": 7253740916505.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916591.199, "dur": 5.332,
    "args": {
      "External id": 1413,      "device": 2, "stream": 0,      "correlation": 715, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 715, "pid": 2, "tid": 0, "ts": 7253740916591.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916596.532, "dur": 5.972,
    "args": {
      "External id": 1414,      "device": 2, "stream": 0,      "correlation": 716, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 716, "pid": 2, "tid": 0, "ts": 7253740916596.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917433.503, "dur": 4.944,
    "args": {
      "External id": 2893,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1455,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1455, "pid": 49625, "tid": 49625, "ts": 7253740917433.503,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917444.069, "dur": 3.576,
    "args": {
      "External id": 2894,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1456,      "grid": [4, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1456, "pid": 49625, "tid": 49625, "ts": 7253740917444.069,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917452.365, "dur": 3.802,
    "args": {
      "External id": 2895,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1457,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1457, "pid": 49625, "tid": 49625, "ts": 7253740917452.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917459.850, "dur": 3.889,
    "args": {
      "External id": 2896,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1458,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1458, "pid": 49625, "tid": 49625, "ts": 7253740917459.850,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917468.112, "dur": 3.576,
    "args": {
      "External id": 2897,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1459,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1459, "pid": 49625, "tid": 49625, "ts": 7253740917468.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917477.083, "dur": 4.286,
    "args": {
      "External id": 2898,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1460,      "grid": [32768, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1460, "pid": 49625, "tid": 49625, "ts": 7253740917477.083,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917485.652, "dur": 3.475,
    "args": {
      "External id": 2900,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 1461,      "grid": [65536, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1461, "pid": 49625, "tid": 49625, "ts": 7253740917485.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 49625, "tid": 49625,
    "ts": 7253740917492.440, "dur": 3.537,
    "args": {
      "External id": 2901,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1462,      "grid": [131072, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1462, "pid": 49625, "tid": 49625, "ts": 7253740917492.440,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740916602.505, "dur": 82.389,
    "args": {
      "External id": 1415,      "device": 2, "stream": 0,      "correlation": 717, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 717, "pid": 2, "tid": 0, "ts": 7253740916602.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916684.895, "dur": 60.217,
    "args": {
      "External id": 1416,      "device": 2, "stream": 0,      "correlation": 718, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 718, "pid": 2, "tid": 0, "ts": 7253740916684.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740916745.113, "dur": 76.374,
    "args": {
      "External id": 1417,      "device": 2, "stream": 0,      "correlation": 719, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 719, "pid": 2, "tid": 0, "ts": 7253740916745.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740916821.488, "dur": 65.430,
    "args": {
      "External id": 1418,      "device": 2, "stream": 0,      "correlation": 720, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 720, "pid": 2, "tid": 0, "ts": 7253740916821.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740916886.919, "dur": 74.891,
    "args": {
      "External id": 1420,      "device": 2, "stream": 0,      "correlation": 721, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 721, "pid": 2, "tid": 0, "ts": 7253740916886.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740916961.811, "dur": 80.744,
    "args": {
      "External id": 1421,      "device": 2, "stream": 0,      "correlation": 722, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 722, "pid": 2, "tid": 0, "ts": 7253740916961.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917049.572, "dur": 45.584,
    "args": {
      "External id": 1423,      "device": 2, "stream": 0,      "correlation": 723, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 723, "pid": 2, "tid": 0, "ts": 7253740917049.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740917095.157, "dur": 78.980,
    "args": {
      "External id": 1426,      "device": 2, "stream": 0,      "correlation": 724, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1699.367290867424
    }
  },
  {
    "ph": "f", "id": 724, "pid": 2, "tid": 0, "ts": 7253740917095.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917181.114, "dur": 5.251,
    "args": {
      "External id": 1433,      "device": 2, "stream": 0,      "correlation": 725, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 725, "pid": 2, "tid": 0, "ts": 7253740917181.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917186.366, "dur": 5.773,
    "args": {
      "External id": 1434,      "device": 2, "stream": 0,      "correlation": 726, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 726, "pid": 2, "tid": 0, "ts": 7253740917186.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740917192.140, "dur": 80.544,
    "args": {
      "External id": 1435,      "device": 2, "stream": 0,      "correlation": 727, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 727, "pid": 2, "tid": 0, "ts": 7253740917192.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917272.685, "dur": 60.217,
    "args": {
      "External id": 1436,      "device": 2, "stream": 0,      "correlation": 728, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 728, "pid": 2, "tid": 0, "ts": 7253740917272.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740917332.903, "dur": 75.973,
    "args": {
      "External id": 1437,      "device": 2, "stream": 0,      "correlation": 729, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 729, "pid": 2, "tid": 0, "ts": 7253740917332.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917408.877, "dur": 64.708,
    "args": {
      "External id": 1438,      "device": 2, "stream": 0,      "correlation": 730, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 730, "pid": 2, "tid": 0, "ts": 7253740917408.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740917473.586, "dur": 74.811,
    "args": {
      "External id": 1440,      "device": 2, "stream": 0,      "correlation": 731, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 731, "pid": 2, "tid": 0, "ts": 7253740917473.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740917548.398, "dur": 80.784,
    "args": {
      "External id": 1441,      "device": 2, "stream": 0,      "correlation": 732, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 732, "pid": 2, "tid": 0, "ts": 7253740917548.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917636.039, "dur": 44.782,
    "args": {
      "External id": 1443,      "device": 2, "stream": 0,      "correlation": 733, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 733, "pid": 2, "tid": 0, "ts": 7253740917636.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740917680.822, "dur": 79.381,
    "args": {
      "External id": 1446,      "device": 2, "stream": 0,      "correlation": 734, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1690.7828978861708
    }
  },
  {
    "ph": "f", "id": 734, "pid": 2, "tid": 0, "ts": 7253740917680.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917766.699, "dur": 5.211,
    "args": {
      "External id": 1453,      "device": 2, "stream": 0,      "correlation": 735, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 735, "pid": 2, "tid": 0, "ts": 7253740917766.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917771.911, "dur": 5.532,
    "args": {
      "External id": 1454,      "device": 2, "stream": 0,      "correlation": 736, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 736, "pid": 2, "tid": 0, "ts": 7253740917771.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740917777.444, "dur": 80.784,
    "args": {
      "External id": 1455,      "device": 2, "stream": 0,      "correlation": 737, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 737, "pid": 2, "tid": 0, "ts": 7253740917777.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917858.229, "dur": 63.465,
    "args": {
      "External id": 1456,      "device": 2, "stream": 0,      "correlation": 738, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 738, "pid": 2, "tid": 0, "ts": 7253740917858.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740917921.695, "dur": 75.773,
    "args": {
      "External id": 1457,      "device": 2, "stream": 0,      "correlation": 739, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 739, "pid": 2, "tid": 0, "ts": 7253740917921.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740917997.469, "dur": 62.182,
    "args": {
      "External id": 1458,      "device": 2, "stream": 0,      "correlation": 740, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 740, "pid": 2, "tid": 0, "ts": 7253740917997.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740918059.652, "dur": 75.412,
    "args": {
      "External id": 1460,      "device": 2, "stream": 0,      "correlation": 741, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 741, "pid": 2, "tid": 0, "ts": 7253740918059.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740918135.065, "dur": 81.266,
    "args": {
      "External id": 1461,      "device": 2, "stream": 0,      "correlation": 742, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 742, "pid": 2, "tid": 0, "ts": 7253740918135.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918223.188, "dur": 46.345,
    "args": {
      "External id": 1463,      "device": 2, "stream": 0,      "correlation": 743, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 743, "pid": 2, "tid": 0, "ts": 7253740918223.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740918269.534, "dur": 78.299,
    "args": {
      "External id": 1466,      "device": 2, "stream": 0,      "correlation": 744, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1714.1472286079184
    }
  },
  {
    "ph": "f", "id": 744, "pid": 2, "tid": 0, "ts": 7253740918269.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918355.211, "dur": 5.171,
    "args": {
      "External id": 1473,      "device": 2, "stream": 0,      "correlation": 745, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 745, "pid": 2, "tid": 0, "ts": 7253740918355.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918360.383, "dur": 5.371,
    "args": {
      "External id": 1474,      "device": 2, "stream": 0,      "correlation": 746, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 746, "pid": 2, "tid": 0, "ts": 7253740918360.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740918365.755, "dur": 80.905,
    "args": {
      "External id": 1475,      "device": 2, "stream": 0,      "correlation": 747, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 747, "pid": 2, "tid": 0, "ts": 7253740918365.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918446.661, "dur": 64.347,
    "args": {
      "External id": 1476,      "device": 2, "stream": 0,      "correlation": 748, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 748, "pid": 2, "tid": 0, "ts": 7253740918446.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740918511.009, "dur": 76.134,
    "args": {
      "External id": 1477,      "device": 2, "stream": 0,      "correlation": 749, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 749, "pid": 2, "tid": 0, "ts": 7253740918511.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918587.144, "dur": 65.910,
    "args": {
      "External id": 1478,      "device": 2, "stream": 0,      "correlation": 750, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 750, "pid": 2, "tid": 0, "ts": 7253740918587.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740918653.055, "dur": 74.169,
    "args": {
      "External id": 1480,      "device": 2, "stream": 0,      "correlation": 751, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 751, "pid": 2, "tid": 0, "ts": 7253740918653.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740918727.225, "dur": 80.424,
    "args": {
      "External id": 1481,      "device": 2, "stream": 0,      "correlation": 752, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 752, "pid": 2, "tid": 0, "ts": 7253740918727.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918814.666, "dur": 45.022,
    "args": {
      "External id": 1483,      "device": 2, "stream": 0,      "correlation": 753, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 753, "pid": 2, "tid": 0, "ts": 7253740918814.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740918859.689, "dur": 77.297,
    "args": {
      "External id": 1486,      "device": 2, "stream": 0,      "correlation": 754, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1736.3674092473286
    }
  },
  {
    "ph": "f", "id": 754, "pid": 2, "tid": 0, "ts": 7253740918859.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918944.404, "dur": 5.331,
    "args": {
      "External id": 1493,      "device": 2, "stream": 0,      "correlation": 755, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 755, "pid": 2, "tid": 0, "ts": 7253740918944.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740918949.736, "dur": 5.893,
    "args": {
      "External id": 1494,      "device": 2, "stream": 0,      "correlation": 756, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 756, "pid": 2, "tid": 0, "ts": 7253740918949.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740918955.630, "dur": 84.192,
    "args": {
      "External id": 1495,      "device": 2, "stream": 0,      "correlation": 757, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 757, "pid": 2, "tid": 0, "ts": 7253740918955.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919039.823, "dur": 60.338,
    "args": {
      "External id": 1496,      "device": 2, "stream": 0,      "correlation": 758, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 758, "pid": 2, "tid": 0, "ts": 7253740919039.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740919100.162, "dur": 79.141,
    "args": {
      "External id": 1497,      "device": 2, "stream": 0,      "correlation": 759, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 759, "pid": 2, "tid": 0, "ts": 7253740919100.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919179.304, "dur": 62.302,
    "args": {
      "External id": 1498,      "device": 2, "stream": 0,      "correlation": 760, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 760, "pid": 2, "tid": 0, "ts": 7253740919179.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740919241.607, "dur": 77.857,
    "args": {
      "External id": 1500,      "device": 2, "stream": 0,      "correlation": 761, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 761, "pid": 2, "tid": 0, "ts": 7253740919241.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740919319.465, "dur": 84.313,
    "args": {
      "External id": 1501,      "device": 2, "stream": 0,      "correlation": 762, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 762, "pid": 2, "tid": 0, "ts": 7253740919319.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919411.397, "dur": 45.784,
    "args": {
      "External id": 1503,      "device": 2, "stream": 0,      "correlation": 763, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 763, "pid": 2, "tid": 0, "ts": 7253740919411.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740919457.182, "dur": 76.053,
    "args": {
      "External id": 1506,      "device": 2, "stream": 0,      "correlation": 764, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1764.768822152681
    }
  },
  {
    "ph": "f", "id": 764, "pid": 2, "tid": 0, "ts": 7253740919457.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919540.253, "dur": 5.331,
    "args": {
      "External id": 1513,      "device": 2, "stream": 0,      "correlation": 765, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 765, "pid": 2, "tid": 0, "ts": 7253740919540.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919545.585, "dur": 5.411,
    "args": {
      "External id": 1514,      "device": 2, "stream": 0,      "correlation": 766, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 766, "pid": 2, "tid": 0, "ts": 7253740919545.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740919550.997, "dur": 84.112,
    "args": {
      "External id": 1515,      "device": 2, "stream": 0,      "correlation": 767, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 767, "pid": 2, "tid": 0, "ts": 7253740919550.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919635.110, "dur": 62.623,
    "args": {
      "External id": 1516,      "device": 2, "stream": 0,      "correlation": 768, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 768, "pid": 2, "tid": 0, "ts": 7253740919635.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740919697.734, "dur": 78.179,
    "args": {
      "External id": 1517,      "device": 2, "stream": 0,      "correlation": 769, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 769, "pid": 2, "tid": 0, "ts": 7253740919697.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740919775.914, "dur": 63.465,
    "args": {
      "External id": 1518,      "device": 2, "stream": 0,      "correlation": 770, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 770, "pid": 2, "tid": 0, "ts": 7253740919775.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740919839.380, "dur": 77.537,
    "args": {
      "External id": 1520,      "device": 2, "stream": 0,      "correlation": 771, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 771, "pid": 2, "tid": 0, "ts": 7253740919839.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740919916.918, "dur": 83.110,
    "args": {
      "External id": 1521,      "device": 2, "stream": 0,      "correlation": 772, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 772, "pid": 2, "tid": 0, "ts": 7253740919916.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920006.764, "dur": 44.581,
    "args": {
      "External id": 1523,      "device": 2, "stream": 0,      "correlation": 773, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 773, "pid": 2, "tid": 0, "ts": 7253740920006.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740920051.346, "dur": 78.259,
    "args": {
      "External id": 1526,      "device": 2, "stream": 0,      "correlation": 774, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1715.0233580373115
    }
  },
  {
    "ph": "f", "id": 774, "pid": 2, "tid": 0, "ts": 7253740920051.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920136.943, "dur": 5.411,
    "args": {
      "External id": 1533,      "device": 2, "stream": 0,      "correlation": 775, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 775, "pid": 2, "tid": 0, "ts": 7253740920136.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920142.355, "dur": 6.013,
    "args": {
      "External id": 1534,      "device": 2, "stream": 0,      "correlation": 776, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 776, "pid": 2, "tid": 0, "ts": 7253740920142.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740920148.369, "dur": 83.872,
    "args": {
      "External id": 1535,      "device": 2, "stream": 0,      "correlation": 777, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 777, "pid": 2, "tid": 0, "ts": 7253740920148.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920232.242, "dur": 61.580,
    "args": {
      "External id": 1536,      "device": 2, "stream": 0,      "correlation": 778, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 778, "pid": 2, "tid": 0, "ts": 7253740920232.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740920293.823, "dur": 78.660,
    "args": {
      "External id": 1537,      "device": 2, "stream": 0,      "correlation": 779, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 779, "pid": 2, "tid": 0, "ts": 7253740920293.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920372.484, "dur": 63.906,
    "args": {
      "External id": 1538,      "device": 2, "stream": 0,      "correlation": 780, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 780, "pid": 2, "tid": 0, "ts": 7253740920372.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740920436.391, "dur": 76.735,
    "args": {
      "External id": 1540,      "device": 2, "stream": 0,      "correlation": 781, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 781, "pid": 2, "tid": 0, "ts": 7253740920436.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740920513.127, "dur": 84.192,
    "args": {
      "External id": 1541,      "device": 2, "stream": 0,      "correlation": 782, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 782, "pid": 2, "tid": 0, "ts": 7253740920513.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920604.296, "dur": 44.903,
    "args": {
      "External id": 1543,      "device": 2, "stream": 0,      "correlation": 783, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 783, "pid": 2, "tid": 0, "ts": 7253740920604.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740920649.200, "dur": 78.940,
    "args": {
      "External id": 1546,      "device": 2, "stream": 0,      "correlation": 784, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1700.2283730887625
    }
  },
  {
    "ph": "f", "id": 784, "pid": 2, "tid": 0, "ts": 7253740920649.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920735.438, "dur": 5.371,
    "args": {
      "External id": 1553,      "device": 2, "stream": 0,      "correlation": 785, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 785, "pid": 2, "tid": 0, "ts": 7253740920735.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920740.810, "dur": 5.572,
    "args": {
      "External id": 1554,      "device": 2, "stream": 0,      "correlation": 786, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 786, "pid": 2, "tid": 0, "ts": 7253740920740.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740920746.383, "dur": 83.831,
    "args": {
      "External id": 1555,      "device": 2, "stream": 0,      "correlation": 787, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 787, "pid": 2, "tid": 0, "ts": 7253740920746.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920830.215, "dur": 60.819,
    "args": {
      "External id": 1556,      "device": 2, "stream": 0,      "correlation": 788, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 788, "pid": 2, "tid": 0, "ts": 7253740920830.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740920891.035, "dur": 78.620,
    "args": {
      "External id": 1557,      "device": 2, "stream": 0,      "correlation": 789, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 789, "pid": 2, "tid": 0, "ts": 7253740920891.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740920969.656, "dur": 64.346,
    "args": {
      "External id": 1558,      "device": 2, "stream": 0,      "correlation": 790, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 790, "pid": 2, "tid": 0, "ts": 7253740920969.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740921034.003, "dur": 77.056,
    "args": {
      "External id": 1560,      "device": 2, "stream": 0,      "correlation": 791, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 791, "pid": 2, "tid": 0, "ts": 7253740921034.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740921111.060, "dur": 84.473,
    "args": {
      "External id": 1561,      "device": 2, "stream": 0,      "correlation": 792, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 792, "pid": 2, "tid": 0, "ts": 7253740921111.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921202.510, "dur": 44.341,
    "args": {
      "External id": 1563,      "device": 2, "stream": 0,      "correlation": 793, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 793, "pid": 2, "tid": 0, "ts": 7253740921202.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740921246.852, "dur": 79.101,
    "args": {
      "External id": 1566,      "device": 2, "stream": 0,      "correlation": 794, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1696.7678187656443
    }
  },
  {
    "ph": "f", "id": 794, "pid": 2, "tid": 0, "ts": 7253740921246.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921333.411, "dur": 5.331,
    "args": {
      "External id": 1573,      "device": 2, "stream": 0,      "correlation": 795, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 795, "pid": 2, "tid": 0, "ts": 7253740921333.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921338.743, "dur": 5.772,
    "args": {
      "External id": 1574,      "device": 2, "stream": 0,      "correlation": 796, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 796, "pid": 2, "tid": 0, "ts": 7253740921338.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740921344.516, "dur": 83.992,
    "args": {
      "External id": 1575,      "device": 2, "stream": 0,      "correlation": 797, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 797, "pid": 2, "tid": 0, "ts": 7253740921344.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921428.509, "dur": 58.895,
    "args": {
      "External id": 1576,      "device": 2, "stream": 0,      "correlation": 798, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 798, "pid": 2, "tid": 0, "ts": 7253740921428.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740921494.461, "dur": 77.737,
    "args": {
      "External id": 1577,      "device": 2, "stream": 0,      "correlation": 799, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 799, "pid": 2, "tid": 0, "ts": 7253740921494.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921572.199, "dur": 62.744,
    "args": {
      "External id": 1578,      "device": 2, "stream": 0,      "correlation": 800, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 800, "pid": 2, "tid": 0, "ts": 7253740921572.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740921634.944, "dur": 76.895,
    "args": {
      "External id": 1580,      "device": 2, "stream": 0,      "correlation": 801, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 801, "pid": 2, "tid": 0, "ts": 7253740921634.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740921711.840, "dur": 84.593,
    "args": {
      "External id": 1581,      "device": 2, "stream": 0,      "correlation": 802, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 802, "pid": 2, "tid": 0, "ts": 7253740921711.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921803.571, "dur": 44.782,
    "args": {
      "External id": 1583,      "device": 2, "stream": 0,      "correlation": 803, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 803, "pid": 2, "tid": 0, "ts": 7253740921803.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740921848.354, "dur": 81.025,
    "args": {
      "External id": 1586,      "device": 2, "stream": 0,      "correlation": 804, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1656.4772788981315
    }
  },
  {
    "ph": "f", "id": 804, "pid": 2, "tid": 0, "ts": 7253740921848.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921936.837, "dur": 5.331,
    "args": {
      "External id": 1593,      "device": 2, "stream": 0,      "correlation": 805, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 805, "pid": 2, "tid": 0, "ts": 7253740921936.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740921942.169, "dur": 5.973,
    "args": {
      "External id": 1594,      "device": 2, "stream": 0,      "correlation": 806, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 806, "pid": 2, "tid": 0, "ts": 7253740921942.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740921948.143, "dur": 82.629,
    "args": {
      "External id": 1595,      "device": 2, "stream": 0,      "correlation": 807, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 807, "pid": 2, "tid": 0, "ts": 7253740921948.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922030.773, "dur": 58.974,
    "args": {
      "External id": 1596,      "device": 2, "stream": 0,      "correlation": 808, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 808, "pid": 2, "tid": 0, "ts": 7253740922030.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740922089.748, "dur": 76.936,
    "args": {
      "External id": 1597,      "device": 2, "stream": 0,      "correlation": 809, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 809, "pid": 2, "tid": 0, "ts": 7253740922089.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922166.685, "dur": 60.458,
    "args": {
      "External id": 1598,      "device": 2, "stream": 0,      "correlation": 810, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 810, "pid": 2, "tid": 0, "ts": 7253740922166.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740922227.144, "dur": 75.131,
    "args": {
      "External id": 1600,      "device": 2, "stream": 0,      "correlation": 811, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 811, "pid": 2, "tid": 0, "ts": 7253740922227.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740922302.276, "dur": 82.308,
    "args": {
      "External id": 1601,      "device": 2, "stream": 0,      "correlation": 812, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 812, "pid": 2, "tid": 0, "ts": 7253740922302.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922391.682, "dur": 45.984,
    "args": {
      "External id": 1603,      "device": 2, "stream": 0,      "correlation": 813, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 813, "pid": 2, "tid": 0, "ts": 7253740922391.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740922437.667, "dur": 77.136,
    "args": {
      "External id": 1606,      "device": 2, "stream": 0,      "correlation": 814, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1739.9915475063847
    }
  },
  {
    "ph": "f", "id": 814, "pid": 2, "tid": 0, "ts": 7253740922437.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922521.820, "dur": 5.412,
    "args": {
      "External id": 1613,      "device": 2, "stream": 0,      "correlation": 815, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 815, "pid": 2, "tid": 0, "ts": 7253740922521.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922527.233, "dur": 5.933,
    "args": {
      "External id": 1614,      "device": 2, "stream": 0,      "correlation": 816, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 816, "pid": 2, "tid": 0, "ts": 7253740922527.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740922533.167, "dur": 82.789,
    "args": {
      "External id": 1615,      "device": 2, "stream": 0,      "correlation": 817, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 817, "pid": 2, "tid": 0, "ts": 7253740922533.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922615.957, "dur": 59.535,
    "args": {
      "External id": 1616,      "device": 2, "stream": 0,      "correlation": 818, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 818, "pid": 2, "tid": 0, "ts": 7253740922615.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740922675.493, "dur": 77.578,
    "args": {
      "External id": 1617,      "device": 2, "stream": 0,      "correlation": 819, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 819, "pid": 2, "tid": 0, "ts": 7253740922675.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922753.072, "dur": 64.507,
    "args": {
      "External id": 1618,      "device": 2, "stream": 0,      "correlation": 820, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 820, "pid": 2, "tid": 0, "ts": 7253740922753.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740922817.580, "dur": 75.131,
    "args": {
      "External id": 1620,      "device": 2, "stream": 0,      "correlation": 821, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 821, "pid": 2, "tid": 0, "ts": 7253740922817.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740922892.712, "dur": 80.985,
    "args": {
      "External id": 1621,      "device": 2, "stream": 0,      "correlation": 822, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 822, "pid": 2, "tid": 0, "ts": 7253740922892.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740922980.594, "dur": 45.744,
    "args": {
      "External id": 1623,      "device": 2, "stream": 0,      "correlation": 823, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 823, "pid": 2, "tid": 0, "ts": 7253740922980.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740923026.339, "dur": 76.695,
    "args": {
      "External id": 1626,      "device": 2, "stream": 0,      "correlation": 824, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1749.996453530823
    }
  },
  {
    "ph": "f", "id": 824, "pid": 2, "tid": 0, "ts": 7253740923026.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923109.891, "dur": 5.412,
    "args": {
      "External id": 1633,      "device": 2, "stream": 0,      "correlation": 825, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 825, "pid": 2, "tid": 0, "ts": 7253740923109.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923115.304, "dur": 5.411,
    "args": {
      "External id": 1634,      "device": 2, "stream": 0,      "correlation": 826, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 826, "pid": 2, "tid": 0, "ts": 7253740923115.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740923120.716, "dur": 80.825,
    "args": {
      "External id": 1635,      "device": 2, "stream": 0,      "correlation": 827, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 827, "pid": 2, "tid": 0, "ts": 7253740923120.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923201.542, "dur": 64.787,
    "args": {
      "External id": 1636,      "device": 2, "stream": 0,      "correlation": 828, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 828, "pid": 2, "tid": 0, "ts": 7253740923201.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740923266.330, "dur": 76.335,
    "args": {
      "External id": 1637,      "device": 2, "stream": 0,      "correlation": 829, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 829, "pid": 2, "tid": 0, "ts": 7253740923266.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923342.666, "dur": 62.903,
    "args": {
      "External id": 1638,      "device": 2, "stream": 0,      "correlation": 830, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 830, "pid": 2, "tid": 0, "ts": 7253740923342.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740923405.570, "dur": 74.611,
    "args": {
      "External id": 1640,      "device": 2, "stream": 0,      "correlation": 831, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 831, "pid": 2, "tid": 0, "ts": 7253740923405.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740923480.182, "dur": 80.463,
    "args": {
      "External id": 1641,      "device": 2, "stream": 0,      "correlation": 832, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 832, "pid": 2, "tid": 0, "ts": 7253740923480.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923567.622, "dur": 44.983,
    "args": {
      "External id": 1643,      "device": 2, "stream": 0,      "correlation": 833, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 833, "pid": 2, "tid": 0, "ts": 7253740923567.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740923612.606, "dur": 78.018,
    "args": {
      "External id": 1646,      "device": 2, "stream": 0,      "correlation": 834, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1720.321050000641
    }
  },
  {
    "ph": "f", "id": 834, "pid": 2, "tid": 0, "ts": 7253740923612.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923697.561, "dur": 5.131,
    "args": {
      "External id": 1653,      "device": 2, "stream": 0,      "correlation": 835, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 835, "pid": 2, "tid": 0, "ts": 7253740923697.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923702.693, "dur": 5.331,
    "args": {
      "External id": 1654,      "device": 2, "stream": 0,      "correlation": 836, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 836, "pid": 2, "tid": 0, "ts": 7253740923702.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740923708.025, "dur": 81.025,
    "args": {
      "External id": 1655,      "device": 2, "stream": 0,      "correlation": 837, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 837, "pid": 2, "tid": 0, "ts": 7253740923708.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923789.051, "dur": 60.097,
    "args": {
      "External id": 1656,      "device": 2, "stream": 0,      "correlation": 838, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 838, "pid": 2, "tid": 0, "ts": 7253740923789.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740923849.149, "dur": 77.738,
    "args": {
      "External id": 1657,      "device": 2, "stream": 0,      "correlation": 839, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 839, "pid": 2, "tid": 0, "ts": 7253740923849.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740923926.888, "dur": 65.309,
    "args": {
      "External id": 1658,      "device": 2, "stream": 0,      "correlation": 840, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 840, "pid": 2, "tid": 0, "ts": 7253740923926.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740923992.198, "dur": 74.089,
    "args": {
      "External id": 1660,      "device": 2, "stream": 0,      "correlation": 841, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 841, "pid": 2, "tid": 0, "ts": 7253740923992.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740924066.288, "dur": 80.945,
    "args": {
      "External id": 1661,      "device": 2, "stream": 0,      "correlation": 842, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 842, "pid": 2, "tid": 0, "ts": 7253740924066.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924154.330, "dur": 45.062,
    "args": {
      "External id": 1663,      "device": 2, "stream": 0,      "correlation": 843, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 843, "pid": 2, "tid": 0, "ts": 7253740924154.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740924199.393, "dur": 77.938,
    "args": {
      "External id": 1666,      "device": 2, "stream": 0,      "correlation": 844, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1722.0868628029614
    }
  },
  {
    "ph": "f", "id": 844, "pid": 2, "tid": 0, "ts": 7253740924199.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924284.228, "dur": 5.332,
    "args": {
      "External id": 1673,      "device": 2, "stream": 0,      "correlation": 845, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 845, "pid": 2, "tid": 0, "ts": 7253740924284.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924289.561, "dur": 5.692,
    "args": {
      "External id": 1674,      "device": 2, "stream": 0,      "correlation": 846, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 846, "pid": 2, "tid": 0, "ts": 7253740924289.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740924295.254, "dur": 80.463,
    "args": {
      "External id": 1675,      "device": 2, "stream": 0,      "correlation": 847, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 847, "pid": 2, "tid": 0, "ts": 7253740924295.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924375.718, "dur": 61.982,
    "args": {
      "External id": 1676,      "device": 2, "stream": 0,      "correlation": 848, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 848, "pid": 2, "tid": 0, "ts": 7253740924375.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740924437.701, "dur": 75.372,
    "args": {
      "External id": 1677,      "device": 2, "stream": 0,      "correlation": 849, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 849, "pid": 2, "tid": 0, "ts": 7253740924437.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924513.074, "dur": 64.427,
    "args": {
      "External id": 1678,      "device": 2, "stream": 0,      "correlation": 850, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 850, "pid": 2, "tid": 0, "ts": 7253740924513.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740924577.502, "dur": 73.568,
    "args": {
      "External id": 1680,      "device": 2, "stream": 0,      "correlation": 851, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 851, "pid": 2, "tid": 0, "ts": 7253740924577.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740924651.071, "dur": 79.742,
    "args": {
      "External id": 1681,      "device": 2, "stream": 0,      "correlation": 852, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 852, "pid": 2, "tid": 0, "ts": 7253740924651.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924737.870, "dur": 45.343,
    "args": {
      "External id": 1683,      "device": 2, "stream": 0,      "correlation": 853, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 853, "pid": 2, "tid": 0, "ts": 7253740924737.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740924783.214, "dur": 79.342,
    "args": {
      "External id": 1686,      "device": 2, "stream": 0,      "correlation": 854, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1691.6139798091829
    }
  },
  {
    "ph": "f", "id": 854, "pid": 2, "tid": 0, "ts": 7253740924783.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924869.492, "dur": 5.332,
    "args": {
      "External id": 1693,      "device": 2, "stream": 0,      "correlation": 855, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 855, "pid": 2, "tid": 0, "ts": 7253740924869.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924874.825, "dur": 5.772,
    "args": {
      "External id": 1694,      "device": 2, "stream": 0,      "correlation": 856, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 856, "pid": 2, "tid": 0, "ts": 7253740924874.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740924880.598, "dur": 80.825,
    "args": {
      "External id": 1695,      "device": 2, "stream": 0,      "correlation": 857, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 857, "pid": 2, "tid": 0, "ts": 7253740924880.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740924961.424, "dur": 63.745,
    "args": {
      "External id": 1696,      "device": 2, "stream": 0,      "correlation": 858, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 858, "pid": 2, "tid": 0, "ts": 7253740924961.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740925025.170, "dur": 76.334,
    "args": {
      "External id": 1697,      "device": 2, "stream": 0,      "correlation": 859, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 859, "pid": 2, "tid": 0, "ts": 7253740925025.170,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925101.505, "dur": 62.824,
    "args": {
      "External id": 1698,      "device": 2, "stream": 0,      "correlation": 860, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 860, "pid": 2, "tid": 0, "ts": 7253740925101.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740925164.330, "dur": 75.332,
    "args": {
      "External id": 1700,      "device": 2, "stream": 0,      "correlation": 861, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 861, "pid": 2, "tid": 0, "ts": 7253740925164.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740925239.663, "dur": 80.864,
    "args": {
      "External id": 1701,      "device": 2, "stream": 0,      "correlation": 862, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 862, "pid": 2, "tid": 0, "ts": 7253740925239.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925327.705, "dur": 44.782,
    "args": {
      "External id": 1703,      "device": 2, "stream": 0,      "correlation": 863, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 863, "pid": 2, "tid": 0, "ts": 7253740925327.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740925372.488, "dur": 79.341,
    "args": {
      "External id": 1706,      "device": 2, "stream": 0,      "correlation": 864, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1691.6353003453405
    }
  },
  {
    "ph": "f", "id": 864, "pid": 2, "tid": 0, "ts": 7253740925372.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925458.846, "dur": 5.331,
    "args": {
      "External id": 1713,      "device": 2, "stream": 0,      "correlation": 865, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 865, "pid": 2, "tid": 0, "ts": 7253740925458.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925464.178, "dur": 5.893,
    "args": {
      "External id": 1714,      "device": 2, "stream": 0,      "correlation": 866, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 866, "pid": 2, "tid": 0, "ts": 7253740925464.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740925470.072, "dur": 81.225,
    "args": {
      "External id": 1715,      "device": 2, "stream": 0,      "correlation": 867, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 867, "pid": 2, "tid": 0, "ts": 7253740925470.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925551.298, "dur": 61.140,
    "args": {
      "External id": 1716,      "device": 2, "stream": 0,      "correlation": 868, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 868, "pid": 2, "tid": 0, "ts": 7253740925551.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740925612.439, "dur": 75.612,
    "args": {
      "External id": 1717,      "device": 2, "stream": 0,      "correlation": 869, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 869, "pid": 2, "tid": 0, "ts": 7253740925612.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925688.052, "dur": 63.545,
    "args": {
      "External id": 1718,      "device": 2, "stream": 0,      "correlation": 870, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 870, "pid": 2, "tid": 0, "ts": 7253740925688.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740925751.598, "dur": 74.691,
    "args": {
      "External id": 1720,      "device": 2, "stream": 0,      "correlation": 871, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 871, "pid": 2, "tid": 0, "ts": 7253740925751.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740925826.290, "dur": 83.190,
    "args": {
      "External id": 1721,      "device": 2, "stream": 0,      "correlation": 872, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 872, "pid": 2, "tid": 0, "ts": 7253740925826.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740925916.577, "dur": 45.263,
    "args": {
      "External id": 1723,      "device": 2, "stream": 0,      "correlation": 873, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 873, "pid": 2, "tid": 0, "ts": 7253740925916.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740925961.841, "dur": 75.493,
    "args": {
      "External id": 1726,      "device": 2, "stream": 0,      "correlation": 874, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1777.8595385063713
    }
  },
  {
    "ph": "f", "id": 874, "pid": 2, "tid": 0, "ts": 7253740925961.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926044.471, "dur": 5.331,
    "args": {
      "External id": 1733,      "device": 2, "stream": 0,      "correlation": 875, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 875, "pid": 2, "tid": 0, "ts": 7253740926044.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926049.803, "dur": 5.452,
    "args": {
      "External id": 1734,      "device": 2, "stream": 0,      "correlation": 876, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 876, "pid": 2, "tid": 0, "ts": 7253740926049.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740926055.256, "dur": 85.636,
    "args": {
      "External id": 1735,      "device": 2, "stream": 0,      "correlation": 877, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 877, "pid": 2, "tid": 0, "ts": 7253740926055.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926140.893, "dur": 67.393,
    "args": {
      "External id": 1736,      "device": 2, "stream": 0,      "correlation": 878, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 878, "pid": 2, "tid": 0, "ts": 7253740926140.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740926208.287, "dur": 79.622,
    "args": {
      "External id": 1737,      "device": 2, "stream": 0,      "correlation": 879, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 879, "pid": 2, "tid": 0, "ts": 7253740926208.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926287.910, "dur": 64.588,
    "args": {
      "External id": 1738,      "device": 2, "stream": 0,      "correlation": 880, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 880, "pid": 2, "tid": 0, "ts": 7253740926287.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740926352.499, "dur": 77.537,
    "args": {
      "External id": 1740,      "device": 2, "stream": 0,      "correlation": 881, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 881, "pid": 2, "tid": 0, "ts": 7253740926352.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740926430.037, "dur": 85.515,
    "args": {
      "External id": 1741,      "device": 2, "stream": 0,      "correlation": 882, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 882, "pid": 2, "tid": 0, "ts": 7253740926430.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926523.090, "dur": 45.344,
    "args": {
      "External id": 1743,      "device": 2, "stream": 0,      "correlation": 883, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 883, "pid": 2, "tid": 0, "ts": 7253740926523.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740926568.435, "dur": 78.178,
    "args": {
      "External id": 1746,      "device": 2, "stream": 0,      "correlation": 884, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1716.800266056102
    }
  },
  {
    "ph": "f", "id": 884, "pid": 2, "tid": 0, "ts": 7253740926568.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926653.991, "dur": 5.411,
    "args": {
      "External id": 1753,      "device": 2, "stream": 0,      "correlation": 885, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 885, "pid": 2, "tid": 0, "ts": 7253740926653.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926659.403, "dur": 5.893,
    "args": {
      "External id": 1754,      "device": 2, "stream": 0,      "correlation": 886, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 886, "pid": 2, "tid": 0, "ts": 7253740926659.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740926665.297, "dur": 85.475,
    "args": {
      "External id": 1755,      "device": 2, "stream": 0,      "correlation": 887, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 887, "pid": 2, "tid": 0, "ts": 7253740926665.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926750.773, "dur": 61.260,
    "args": {
      "External id": 1756,      "device": 2, "stream": 0,      "correlation": 888, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 888, "pid": 2, "tid": 0, "ts": 7253740926750.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740926812.034, "dur": 77.697,
    "args": {
      "External id": 1757,      "device": 2, "stream": 0,      "correlation": 889, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 889, "pid": 2, "tid": 0, "ts": 7253740926812.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740926889.732, "dur": 67.394,
    "args": {
      "External id": 1758,      "device": 2, "stream": 0,      "correlation": 890, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 890, "pid": 2, "tid": 0, "ts": 7253740926889.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740926957.127, "dur": 75.853,
    "args": {
      "External id": 1760,      "device": 2, "stream": 0,      "correlation": 891, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 891, "pid": 2, "tid": 0, "ts": 7253740926957.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740927032.981, "dur": 82.910,
    "args": {
      "External id": 1761,      "device": 2, "stream": 0,      "correlation": 892, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 892, "pid": 2, "tid": 0, "ts": 7253740927032.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927122.828, "dur": 45.142,
    "args": {
      "External id": 1763,      "device": 2, "stream": 0,      "correlation": 893, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 893, "pid": 2, "tid": 0, "ts": 7253740927122.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740927167.971, "dur": 76.896,
    "args": {
      "External id": 1766,      "device": 2, "stream": 0,      "correlation": 894, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1745.4221621129564
    }
  },
  {
    "ph": "f", "id": 894, "pid": 2, "tid": 0, "ts": 7253740927167.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927252.125, "dur": 5.411,
    "args": {
      "External id": 1773,      "device": 2, "stream": 0,      "correlation": 895, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 895, "pid": 2, "tid": 0, "ts": 7253740927252.125,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927257.537, "dur": 5.412,
    "args": {
      "External id": 1774,      "device": 2, "stream": 0,      "correlation": 896, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 896, "pid": 2, "tid": 0, "ts": 7253740927257.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740927262.950, "dur": 83.591,
    "args": {
      "External id": 1775,      "device": 2, "stream": 0,      "correlation": 897, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 897, "pid": 2, "tid": 0, "ts": 7253740927262.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927346.542, "dur": 62.021,
    "args": {
      "External id": 1776,      "device": 2, "stream": 0,      "correlation": 898, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 898, "pid": 2, "tid": 0, "ts": 7253740927346.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740927408.564, "dur": 78.780,
    "args": {
      "External id": 1777,      "device": 2, "stream": 0,      "correlation": 899, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 899, "pid": 2, "tid": 0, "ts": 7253740927408.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927487.345, "dur": 63.906,
    "args": {
      "External id": 1778,      "device": 2, "stream": 0,      "correlation": 900, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 900, "pid": 2, "tid": 0, "ts": 7253740927487.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740927551.252, "dur": 75.933,
    "args": {
      "External id": 1780,      "device": 2, "stream": 0,      "correlation": 901, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 901, "pid": 2, "tid": 0, "ts": 7253740927551.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740927627.186, "dur": 83.110,
    "args": {
      "External id": 1781,      "device": 2, "stream": 0,      "correlation": 902, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 902, "pid": 2, "tid": 0, "ts": 7253740927627.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927717.313, "dur": 44.381,
    "args": {
      "External id": 1783,      "device": 2, "stream": 0,      "correlation": 903, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 903, "pid": 2, "tid": 0, "ts": 7253740927717.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740927761.695, "dur": 78.459,
    "args": {
      "External id": 1786,      "device": 2, "stream": 0,      "correlation": 904, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1710.6516441498852
    }
  },
  {
    "ph": "f", "id": 904, "pid": 2, "tid": 0, "ts": 7253740927761.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927847.171, "dur": 5.452,
    "args": {
      "External id": 1793,      "device": 2, "stream": 0,      "correlation": 905, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 905, "pid": 2, "tid": 0, "ts": 7253740927847.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927852.624, "dur": 5.451,
    "args": {
      "External id": 1794,      "device": 2, "stream": 0,      "correlation": 906, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 906, "pid": 2, "tid": 0, "ts": 7253740927852.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740927858.076, "dur": 82.990,
    "args": {
      "External id": 1795,      "device": 2, "stream": 0,      "correlation": 907, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 907, "pid": 2, "tid": 0, "ts": 7253740927858.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740927941.067, "dur": 63.144,
    "args": {
      "External id": 1796,      "device": 2, "stream": 0,      "correlation": 908, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 908, "pid": 2, "tid": 0, "ts": 7253740927941.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740928004.212, "dur": 77.337,
    "args": {
      "External id": 1797,      "device": 2, "stream": 0,      "correlation": 909, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 909, "pid": 2, "tid": 0, "ts": 7253740928004.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928081.550, "dur": 66.712,
    "args": {
      "External id": 1798,      "device": 2, "stream": 0,      "correlation": 910, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 910, "pid": 2, "tid": 0, "ts": 7253740928081.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740928148.263, "dur": 74.931,
    "args": {
      "External id": 1800,      "device": 2, "stream": 0,      "correlation": 911, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 911, "pid": 2, "tid": 0, "ts": 7253740928148.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740928223.195, "dur": 82.388,
    "args": {
      "External id": 1801,      "device": 2, "stream": 0,      "correlation": 912, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 912, "pid": 2, "tid": 0, "ts": 7253740928223.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928313.041, "dur": 44.261,
    "args": {
      "External id": 1803,      "device": 2, "stream": 0,      "correlation": 913, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 913, "pid": 2, "tid": 0, "ts": 7253740928313.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740928357.303, "dur": 77.136,
    "args": {
      "External id": 1806,      "device": 2, "stream": 0,      "correlation": 914, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1739.9915475063847
    }
  },
  {
    "ph": "f", "id": 914, "pid": 2, "tid": 0, "ts": 7253740928357.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928441.657, "dur": 5.331,
    "args": {
      "External id": 1813,      "device": 2, "stream": 0,      "correlation": 915, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 915, "pid": 2, "tid": 0, "ts": 7253740928441.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928446.989, "dur": 5.973,
    "args": {
      "External id": 1814,      "device": 2, "stream": 0,      "correlation": 916, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 916, "pid": 2, "tid": 0, "ts": 7253740928446.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740928452.963, "dur": 82.749,
    "args": {
      "External id": 1815,      "device": 2, "stream": 0,      "correlation": 917, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 917, "pid": 2, "tid": 0, "ts": 7253740928452.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928535.713, "dur": 59.255,
    "args": {
      "External id": 1816,      "device": 2, "stream": 0,      "correlation": 918, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 918, "pid": 2, "tid": 0, "ts": 7253740928535.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740928594.969, "dur": 78.379,
    "args": {
      "External id": 1817,      "device": 2, "stream": 0,      "correlation": 919, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 919, "pid": 2, "tid": 0, "ts": 7253740928594.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928673.349, "dur": 65.189,
    "args": {
      "External id": 1818,      "device": 2, "stream": 0,      "correlation": 920, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 920, "pid": 2, "tid": 0, "ts": 7253740928673.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740928738.539, "dur": 76.534,
    "args": {
      "External id": 1820,      "device": 2, "stream": 0,      "correlation": 921, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 921, "pid": 2, "tid": 0, "ts": 7253740928738.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740928815.074, "dur": 82.629,
    "args": {
      "External id": 1821,      "device": 2, "stream": 0,      "correlation": 922, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 922, "pid": 2, "tid": 0, "ts": 7253740928815.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740928904.881, "dur": 45.343,
    "args": {
      "External id": 1823,      "device": 2, "stream": 0,      "correlation": 923, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 923, "pid": 2, "tid": 0, "ts": 7253740928904.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740928950.225, "dur": 77.617,
    "args": {
      "External id": 1826,      "device": 2, "stream": 0,      "correlation": 924, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1729.2087917751037
    }
  },
  {
    "ph": "f", "id": 924, "pid": 2, "tid": 0, "ts": 7253740928950.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929034.899, "dur": 5.452,
    "args": {
      "External id": 1833,      "device": 2, "stream": 0,      "correlation": 925, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 925, "pid": 2, "tid": 0, "ts": 7253740929034.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929040.352, "dur": 6.013,
    "args": {
      "External id": 1834,      "device": 2, "stream": 0,      "correlation": 926, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 926, "pid": 2, "tid": 0, "ts": 7253740929040.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740929046.366, "dur": 82.508,
    "args": {
      "External id": 1835,      "device": 2, "stream": 0,      "correlation": 927, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 927, "pid": 2, "tid": 0, "ts": 7253740929046.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929128.875, "dur": 64.948,
    "args": {
      "External id": 1836,      "device": 2, "stream": 0,      "correlation": 928, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 928, "pid": 2, "tid": 0, "ts": 7253740929128.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740929193.824, "dur": 76.856,
    "args": {
      "External id": 1837,      "device": 2, "stream": 0,      "correlation": 929, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 929, "pid": 2, "tid": 0, "ts": 7253740929193.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929270.681, "dur": 67.995,
    "args": {
      "External id": 1838,      "device": 2, "stream": 0,      "correlation": 930, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 930, "pid": 2, "tid": 0, "ts": 7253740929270.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740929338.677, "dur": 74.811,
    "args": {
      "External id": 1840,      "device": 2, "stream": 0,      "correlation": 931, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 931, "pid": 2, "tid": 0, "ts": 7253740929338.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740929413.489, "dur": 82.669,
    "args": {
      "External id": 1841,      "device": 2, "stream": 0,      "correlation": 932, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 932, "pid": 2, "tid": 0, "ts": 7253740929413.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929503.255, "dur": 44.541,
    "args": {
      "External id": 1843,      "device": 2, "stream": 0,      "correlation": 933, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 933, "pid": 2, "tid": 0, "ts": 7253740929503.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740929547.797, "dur": 77.898,
    "args": {
      "External id": 1846,      "device": 2, "stream": 0,      "correlation": 934, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1722.9711292827892
    }
  },
  {
    "ph": "f", "id": 934, "pid": 2, "tid": 0, "ts": 7253740929547.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929632.913, "dur": 5.532,
    "args": {
      "External id": 1853,      "device": 2, "stream": 0,      "correlation": 935, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 935, "pid": 2, "tid": 0, "ts": 7253740929632.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929638.446, "dur": 5.892,
    "args": {
      "External id": 1854,      "device": 2, "stream": 0,      "correlation": 936, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 936, "pid": 2, "tid": 0, "ts": 7253740929638.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740929644.339, "dur": 82.468,
    "args": {
      "External id": 1855,      "device": 2, "stream": 0,      "correlation": 937, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 937, "pid": 2, "tid": 0, "ts": 7253740929644.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929726.808, "dur": 64.909,
    "args": {
      "External id": 1856,      "device": 2, "stream": 0,      "correlation": 938, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 938, "pid": 2, "tid": 0, "ts": 7253740929726.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740929791.718, "dur": 77.056,
    "args": {
      "External id": 1857,      "device": 2, "stream": 0,      "correlation": 939, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 939, "pid": 2, "tid": 0, "ts": 7253740929791.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740929868.775, "dur": 64.066,
    "args": {
      "External id": 1858,      "device": 2, "stream": 0,      "correlation": 940, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 940, "pid": 2, "tid": 0, "ts": 7253740929868.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740929932.842, "dur": 75.292,
    "args": {
      "External id": 1860,      "device": 2, "stream": 0,      "correlation": 941, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 941, "pid": 2, "tid": 0, "ts": 7253740929932.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740930008.135, "dur": 82.147,
    "args": {
      "External id": 1861,      "device": 2, "stream": 0,      "correlation": 942, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 942, "pid": 2, "tid": 0, "ts": 7253740930008.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930097.460, "dur": 44.942,
    "args": {
      "External id": 1863,      "device": 2, "stream": 0,      "correlation": 943, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 943, "pid": 2, "tid": 0, "ts": 7253740930097.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740930142.403, "dur": 77.697,
    "args": {
      "External id": 1866,      "device": 2, "stream": 0,      "correlation": 944, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1727.4283507941002
    }
  },
  {
    "ph": "f", "id": 944, "pid": 2, "tid": 0, "ts": 7253740930142.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930226.997, "dur": 5.291,
    "args": {
      "External id": 1873,      "device": 2, "stream": 0,      "correlation": 945, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 945, "pid": 2, "tid": 0, "ts": 7253740930226.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930232.289, "dur": 5.933,
    "args": {
      "External id": 1874,      "device": 2, "stream": 0,      "correlation": 946, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 946, "pid": 2, "tid": 0, "ts": 7253740930232.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740930238.223, "dur": 82.148,
    "args": {
      "External id": 1875,      "device": 2, "stream": 0,      "correlation": 947, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 947, "pid": 2, "tid": 0, "ts": 7253740930238.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930320.372, "dur": 62.703,
    "args": {
      "External id": 1876,      "device": 2, "stream": 0,      "correlation": 948, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 948, "pid": 2, "tid": 0, "ts": 7253740930320.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740930383.076, "dur": 78.018,
    "args": {
      "External id": 1877,      "device": 2, "stream": 0,      "correlation": 949, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 949, "pid": 2, "tid": 0, "ts": 7253740930383.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930461.095, "dur": 66.752,
    "args": {
      "External id": 1878,      "device": 2, "stream": 0,      "correlation": 950, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 950, "pid": 2, "tid": 0, "ts": 7253740930461.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740930527.848, "dur": 76.014,
    "args": {
      "External id": 1880,      "device": 2, "stream": 0,      "correlation": 951, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 951, "pid": 2, "tid": 0, "ts": 7253740930527.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740930603.863, "dur": 81.506,
    "args": {
      "External id": 1881,      "device": 2, "stream": 0,      "correlation": 952, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 952, "pid": 2, "tid": 0, "ts": 7253740930603.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930692.466, "dur": 45.384,
    "args": {
      "External id": 1883,      "device": 2, "stream": 0,      "correlation": 953, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 953, "pid": 2, "tid": 0, "ts": 7253740930692.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740930737.851, "dur": 80.463,
    "args": {
      "External id": 1886,      "device": 2, "stream": 0,      "correlation": 954, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1668.0469278186517
    }
  },
  {
    "ph": "f", "id": 954, "pid": 2, "tid": 0, "ts": 7253740930737.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930825.612, "dur": 5.452,
    "args": {
      "External id": 1893,      "device": 2, "stream": 0,      "correlation": 955, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 955, "pid": 2, "tid": 0, "ts": 7253740930825.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930831.065, "dur": 5.732,
    "args": {
      "External id": 1894,      "device": 2, "stream": 0,      "correlation": 956, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 956, "pid": 2, "tid": 0, "ts": 7253740930831.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740930836.798, "dur": 81.426,
    "args": {
      "External id": 1895,      "device": 2, "stream": 0,      "correlation": 957, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 957, "pid": 2, "tid": 0, "ts": 7253740930836.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740930918.225, "dur": 62.984,
    "args": {
      "External id": 1896,      "device": 2, "stream": 0,      "correlation": 958, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 958, "pid": 2, "tid": 0, "ts": 7253740930918.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740930981.210, "dur": 76.294,
    "args": {
      "External id": 1897,      "device": 2, "stream": 0,      "correlation": 959, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 959, "pid": 2, "tid": 0, "ts": 7253740930981.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931057.505, "dur": 66.191,
    "args": {
      "External id": 1898,      "device": 2, "stream": 0,      "correlation": 960, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 960, "pid": 2, "tid": 0, "ts": 7253740931057.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740931123.697, "dur": 74.971,
    "args": {
      "External id": 1900,      "device": 2, "stream": 0,      "correlation": 961, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 961, "pid": 2, "tid": 0, "ts": 7253740931123.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740931198.669, "dur": 80.263,
    "args": {
      "External id": 1901,      "device": 2, "stream": 0,      "correlation": 962, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 962, "pid": 2, "tid": 0, "ts": 7253740931198.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931285.990, "dur": 45.423,
    "args": {
      "External id": 1903,      "device": 2, "stream": 0,      "correlation": 963, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 963, "pid": 2, "tid": 0, "ts": 7253740931285.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740931331.414, "dur": 79.862,
    "args": {
      "External id": 1906,      "device": 2, "stream": 0,      "correlation": 964, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1680.5996268609995
    }
  },
  {
    "ph": "f", "id": 964, "pid": 2, "tid": 0, "ts": 7253740931331.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931418.213, "dur": 5.291,
    "args": {
      "External id": 1913,      "device": 2, "stream": 0,      "correlation": 965, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 965, "pid": 2, "tid": 0, "ts": 7253740931418.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931423.505, "dur": 6.093,
    "args": {
      "External id": 1914,      "device": 2, "stream": 0,      "correlation": 966, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 966, "pid": 2, "tid": 0, "ts": 7253740931423.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740931429.599, "dur": 80.544,
    "args": {
      "External id": 1915,      "device": 2, "stream": 0,      "correlation": 967, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 967, "pid": 2, "tid": 0, "ts": 7253740931429.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931510.144, "dur": 61.902,
    "args": {
      "External id": 1916,      "device": 2, "stream": 0,      "correlation": 968, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 968, "pid": 2, "tid": 0, "ts": 7253740931510.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740931572.047, "dur": 75.973,
    "args": {
      "External id": 1917,      "device": 2, "stream": 0,      "correlation": 969, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 969, "pid": 2, "tid": 0, "ts": 7253740931572.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931648.021, "dur": 64.467,
    "args": {
      "External id": 1918,      "device": 2, "stream": 0,      "correlation": 970, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 970, "pid": 2, "tid": 0, "ts": 7253740931648.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740931712.489, "dur": 74.009,
    "args": {
      "External id": 1920,      "device": 2, "stream": 0,      "correlation": 971, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 971, "pid": 2, "tid": 0, "ts": 7253740931712.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740931786.499, "dur": 80.263,
    "args": {
      "External id": 1921,      "device": 2, "stream": 0,      "correlation": 972, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 972, "pid": 2, "tid": 0, "ts": 7253740931786.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740931874.140, "dur": 44.461,
    "args": {
      "External id": 1923,      "device": 2, "stream": 0,      "correlation": 973, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 973, "pid": 2, "tid": 0, "ts": 7253740931874.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740931918.602, "dur": 77.097,
    "args": {
      "External id": 1926,      "device": 2, "stream": 0,      "correlation": 974, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1740.8717217048431
    }
  },
  {
    "ph": "f", "id": 974, "pid": 2, "tid": 0, "ts": 7253740931918.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932002.595, "dur": 5.372,
    "args": {
      "External id": 1933,      "device": 2, "stream": 0,      "correlation": 975, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 975, "pid": 2, "tid": 0, "ts": 7253740932002.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932007.968, "dur": 6.013,
    "args": {
      "External id": 1934,      "device": 2, "stream": 0,      "correlation": 976, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 976, "pid": 2, "tid": 0, "ts": 7253740932007.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740932013.982, "dur": 81.867,
    "args": {
      "External id": 1935,      "device": 2, "stream": 0,      "correlation": 977, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 977, "pid": 2, "tid": 0, "ts": 7253740932013.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932095.850, "dur": 65.228,
    "args": {
      "External id": 1936,      "device": 2, "stream": 0,      "correlation": 978, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 978, "pid": 2, "tid": 0, "ts": 7253740932095.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740932161.079, "dur": 77.898,
    "args": {
      "External id": 1937,      "device": 2, "stream": 0,      "correlation": 979, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 979, "pid": 2, "tid": 0, "ts": 7253740932161.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932238.978, "dur": 65.590,
    "args": {
      "External id": 1938,      "device": 2, "stream": 0,      "correlation": 980, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 980, "pid": 2, "tid": 0, "ts": 7253740932238.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740932304.569, "dur": 75.051,
    "args": {
      "External id": 1940,      "device": 2, "stream": 0,      "correlation": 981, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 981, "pid": 2, "tid": 0, "ts": 7253740932304.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740932379.621, "dur": 81.226,
    "args": {
      "External id": 1941,      "device": 2, "stream": 0,      "correlation": 982, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 982, "pid": 2, "tid": 0, "ts": 7253740932379.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932467.744, "dur": 43.859,
    "args": {
      "External id": 1943,      "device": 2, "stream": 0,      "correlation": 983, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 983, "pid": 2, "tid": 0, "ts": 7253740932467.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740932511.604, "dur": 77.538,
    "args": {
      "External id": 1946,      "device": 2, "stream": 0,      "correlation": 984, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1730.9705825455578
    }
  },
  {
    "ph": "f", "id": 984, "pid": 2, "tid": 0, "ts": 7253740932511.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932596.159, "dur": 5.491,
    "args": {
      "External id": 1953,      "device": 2, "stream": 0,      "correlation": 985, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 985, "pid": 2, "tid": 0, "ts": 7253740932596.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932601.651, "dur": 5.813,
    "args": {
      "External id": 1954,      "device": 2, "stream": 0,      "correlation": 986, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 986, "pid": 2, "tid": 0, "ts": 7253740932601.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740932607.465, "dur": 81.867,
    "args": {
      "External id": 1955,      "device": 2, "stream": 0,      "correlation": 987, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 987, "pid": 2, "tid": 0, "ts": 7253740932607.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932689.333, "dur": 65.669,
    "args": {
      "External id": 1956,      "device": 2, "stream": 0,      "correlation": 988, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 988, "pid": 2, "tid": 0, "ts": 7253740932689.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740932755.003, "dur": 77.337,
    "args": {
      "External id": 1957,      "device": 2, "stream": 0,      "correlation": 989, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 989, "pid": 2, "tid": 0, "ts": 7253740932755.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740932832.341, "dur": 65.670,
    "args": {
      "External id": 1958,      "device": 2, "stream": 0,      "correlation": 990, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 990, "pid": 2, "tid": 0, "ts": 7253740932832.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740932898.012, "dur": 76.414,
    "args": {
      "External id": 1960,      "device": 2, "stream": 0,      "correlation": 991, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 991, "pid": 2, "tid": 0, "ts": 7253740932898.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740932974.427, "dur": 82.950,
    "args": {
      "External id": 1961,      "device": 2, "stream": 0,      "correlation": 992, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 992, "pid": 2, "tid": 0, "ts": 7253740932974.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933064.554, "dur": 45.664,
    "args": {
      "External id": 1963,      "device": 2, "stream": 0,      "correlation": 993, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 993, "pid": 2, "tid": 0, "ts": 7253740933064.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740933110.219, "dur": 75.453,
    "args": {
      "External id": 1966,      "device": 2, "stream": 0,      "correlation": 994, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1778.80202507488
    }
  },
  {
    "ph": "f", "id": 994, "pid": 2, "tid": 0, "ts": 7253740933110.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933192.809, "dur": 5.331,
    "args": {
      "External id": 1973,      "device": 2, "stream": 0,      "correlation": 995, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 995, "pid": 2, "tid": 0, "ts": 7253740933192.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933198.141, "dur": 5.732,
    "args": {
      "External id": 1974,      "device": 2, "stream": 0,      "correlation": 996, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 996, "pid": 2, "tid": 0, "ts": 7253740933198.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740933203.874, "dur": 83.231,
    "args": {
      "External id": 1975,      "device": 2, "stream": 0,      "correlation": 997, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 997, "pid": 2, "tid": 0, "ts": 7253740933203.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933287.106, "dur": 60.297,
    "args": {
      "External id": 1976,      "device": 2, "stream": 0,      "correlation": 998, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 998, "pid": 2, "tid": 0, "ts": 7253740933287.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740933347.404, "dur": 77.978,
    "args": {
      "External id": 1977,      "device": 2, "stream": 0,      "correlation": 999, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 999, "pid": 2, "tid": 0, "ts": 7253740933347.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933425.383, "dur": 62.944,
    "args": {
      "External id": 1978,      "device": 2, "stream": 0,      "correlation": 1000, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1000, "pid": 2, "tid": 0, "ts": 7253740933425.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740933488.328, "dur": 76.174,
    "args": {
      "External id": 1980,      "device": 2, "stream": 0,      "correlation": 1001, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1001, "pid": 2, "tid": 0, "ts": 7253740933488.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740933564.503, "dur": 83.671,
    "args": {
      "External id": 1981,      "device": 2, "stream": 0,      "correlation": 1002, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1002, "pid": 2, "tid": 0, "ts": 7253740933564.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933654.790, "dur": 45.223,
    "args": {
      "External id": 1983,      "device": 2, "stream": 0,      "correlation": 1003, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1003, "pid": 2, "tid": 0, "ts": 7253740933654.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740933700.014, "dur": 76.815,
    "args": {
      "External id": 1986,      "device": 2, "stream": 0,      "correlation": 1004, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1747.2626536138305
    }
  },
  {
    "ph": "f", "id": 1004, "pid": 2, "tid": 0, "ts": 7253740933700.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933783.766, "dur": 5.492,
    "args": {
      "External id": 1993,      "device": 2, "stream": 0,      "correlation": 1005, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1005, "pid": 2, "tid": 0, "ts": 7253740933783.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933789.259, "dur": 5.812,
    "args": {
      "External id": 1994,      "device": 2, "stream": 0,      "correlation": 1006, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1006, "pid": 2, "tid": 0, "ts": 7253740933789.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740933795.072, "dur": 82.709,
    "args": {
      "External id": 1995,      "device": 2, "stream": 0,      "correlation": 1007, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1007, "pid": 2, "tid": 0, "ts": 7253740933795.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740933877.782, "dur": 60.298,
    "args": {
      "External id": 1996,      "device": 2, "stream": 0,      "correlation": 1008, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1008, "pid": 2, "tid": 0, "ts": 7253740933877.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740933938.081, "dur": 77.096,
    "args": {
      "External id": 1997,      "device": 2, "stream": 0,      "correlation": 1009, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1009, "pid": 2, "tid": 0, "ts": 7253740933938.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934015.178, "dur": 63.344,
    "args": {
      "External id": 1998,      "device": 2, "stream": 0,      "correlation": 1010, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1010, "pid": 2, "tid": 0, "ts": 7253740934015.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740934078.523, "dur": 74.731,
    "args": {
      "External id": 2000,      "device": 2, "stream": 0,      "correlation": 1011, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1011, "pid": 2, "tid": 0, "ts": 7253740934078.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740934153.255, "dur": 81.346,
    "args": {
      "External id": 2001,      "device": 2, "stream": 0,      "correlation": 1012, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1012, "pid": 2, "tid": 0, "ts": 7253740934153.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934242.099, "dur": 44.862,
    "args": {
      "External id": 2003,      "device": 2, "stream": 0,      "correlation": 1013, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1013, "pid": 2, "tid": 0, "ts": 7253740934242.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740934286.962, "dur": 76.775,
    "args": {
      "External id": 2006,      "device": 2, "stream": 0,      "correlation": 1014, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1748.1729707200166
    }
  },
  {
    "ph": "f", "id": 1014, "pid": 2, "tid": 0, "ts": 7253740934286.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934370.674, "dur": 5.452,
    "args": {
      "External id": 2013,      "device": 2, "stream": 0,      "correlation": 1015, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1015, "pid": 2, "tid": 0, "ts": 7253740934370.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934376.127, "dur": 5.852,
    "args": {
      "External id": 2014,      "device": 2, "stream": 0,      "correlation": 1016, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1016, "pid": 2, "tid": 0, "ts": 7253740934376.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740934381.980, "dur": 81.547,
    "args": {
      "External id": 2015,      "device": 2, "stream": 0,      "correlation": 1017, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1017, "pid": 2, "tid": 0, "ts": 7253740934381.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934463.528, "dur": 62.302,
    "args": {
      "External id": 2016,      "device": 2, "stream": 0,      "correlation": 1018, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1018, "pid": 2, "tid": 0, "ts": 7253740934463.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740934525.831, "dur": 76.414,
    "args": {
      "External id": 2017,      "device": 2, "stream": 0,      "correlation": 1019, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1019, "pid": 2, "tid": 0, "ts": 7253740934525.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934602.246, "dur": 63.264,
    "args": {
      "External id": 2018,      "device": 2, "stream": 0,      "correlation": 1020, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1020, "pid": 2, "tid": 0, "ts": 7253740934602.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740934665.511, "dur": 74.450,
    "args": {
      "External id": 2020,      "device": 2, "stream": 0,      "correlation": 1021, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1021, "pid": 2, "tid": 0, "ts": 7253740934665.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740934739.962, "dur": 81.386,
    "args": {
      "External id": 2021,      "device": 2, "stream": 0,      "correlation": 1022, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1022, "pid": 2, "tid": 0, "ts": 7253740934739.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934829.047, "dur": 44.782,
    "args": {
      "External id": 2023,      "device": 2, "stream": 0,      "correlation": 1023, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1023, "pid": 2, "tid": 0, "ts": 7253740934829.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740934873.830, "dur": 77.858,
    "args": {
      "External id": 2026,      "device": 2, "stream": 0,      "correlation": 1024, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1723.856304345034
    }
  },
  {
    "ph": "f", "id": 1024, "pid": 2, "tid": 0, "ts": 7253740934873.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934958.544, "dur": 5.412,
    "args": {
      "External id": 2033,      "device": 2, "stream": 0,      "correlation": 1025, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1025, "pid": 2, "tid": 0, "ts": 7253740934958.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740934963.957, "dur": 5.772,
    "args": {
      "External id": 2034,      "device": 2, "stream": 0,      "correlation": 1026, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1026, "pid": 2, "tid": 0, "ts": 7253740934963.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740934969.730, "dur": 83.070,
    "args": {
      "External id": 2035,      "device": 2, "stream": 0,      "correlation": 1027, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1027, "pid": 2, "tid": 0, "ts": 7253740934969.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935052.801, "dur": 58.734,
    "args": {
      "External id": 2036,      "device": 2, "stream": 0,      "correlation": 1028, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1028, "pid": 2, "tid": 0, "ts": 7253740935052.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740935111.536, "dur": 77.898,
    "args": {
      "External id": 2037,      "device": 2, "stream": 0,      "correlation": 1029, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1029, "pid": 2, "tid": 0, "ts": 7253740935111.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935189.435, "dur": 66.632,
    "args": {
      "External id": 2038,      "device": 2, "stream": 0,      "correlation": 1030, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1030, "pid": 2, "tid": 0, "ts": 7253740935189.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740935256.068, "dur": 76.053,
    "args": {
      "External id": 2040,      "device": 2, "stream": 0,      "correlation": 1031, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1031, "pid": 2, "tid": 0, "ts": 7253740935256.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740935332.122, "dur": 83.311,
    "args": {
      "External id": 2041,      "device": 2, "stream": 0,      "correlation": 1032, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1032, "pid": 2, "tid": 0, "ts": 7253740935332.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935422.530, "dur": 44.702,
    "args": {
      "External id": 2043,      "device": 2, "stream": 0,      "correlation": 1033, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1033, "pid": 2, "tid": 0, "ts": 7253740935422.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740935467.233, "dur": 78.299,
    "args": {
      "External id": 2046,      "device": 2, "stream": 0,      "correlation": 1034, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1714.1472286079184
    }
  },
  {
    "ph": "f", "id": 1034, "pid": 2, "tid": 0, "ts": 7253740935467.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935552.629, "dur": 5.411,
    "args": {
      "External id": 2053,      "device": 2, "stream": 0,      "correlation": 1035, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1035, "pid": 2, "tid": 0, "ts": 7253740935552.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935558.041, "dur": 5.813,
    "args": {
      "External id": 2054,      "device": 2, "stream": 0,      "correlation": 1036, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1036, "pid": 2, "tid": 0, "ts": 7253740935558.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740935563.855, "dur": 83.110,
    "args": {
      "External id": 2055,      "device": 2, "stream": 0,      "correlation": 1037, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1037, "pid": 2, "tid": 0, "ts": 7253740935563.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935646.966, "dur": 60.016,
    "args": {
      "External id": 2056,      "device": 2, "stream": 0,      "correlation": 1038, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1038, "pid": 2, "tid": 0, "ts": 7253740935646.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740935706.983, "dur": 77.818,
    "args": {
      "External id": 2057,      "device": 2, "stream": 0,      "correlation": 1039, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1039, "pid": 2, "tid": 0, "ts": 7253740935706.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740935784.802, "dur": 63.505,
    "args": {
      "External id": 2058,      "device": 2, "stream": 0,      "correlation": 1040, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1040, "pid": 2, "tid": 0, "ts": 7253740935784.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740935848.308, "dur": 75.853,
    "args": {
      "External id": 2060,      "device": 2, "stream": 0,      "correlation": 1041, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1041, "pid": 2, "tid": 0, "ts": 7253740935848.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740935924.162, "dur": 82.108,
    "args": {
      "External id": 2061,      "device": 2, "stream": 0,      "correlation": 1042, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1042, "pid": 2, "tid": 0, "ts": 7253740935924.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936013.207, "dur": 44.942,
    "args": {
      "External id": 2063,      "device": 2, "stream": 0,      "correlation": 1043, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1043, "pid": 2, "tid": 0, "ts": 7253740936013.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740936058.150, "dur": 78.900,
    "args": {
      "External id": 2066,      "device": 2, "stream": 0,      "correlation": 1044, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1701.0903283862056
    }
  },
  {
    "ph": "f", "id": 1044, "pid": 2, "tid": 0, "ts": 7253740936058.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936144.669, "dur": 5.371,
    "args": {
      "External id": 2073,      "device": 2, "stream": 0,      "correlation": 1045, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1045, "pid": 2, "tid": 0, "ts": 7253740936144.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936150.041, "dur": 5.612,
    "args": {
      "External id": 2074,      "device": 2, "stream": 0,      "correlation": 1046, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1046, "pid": 2, "tid": 0, "ts": 7253740936150.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740936155.654, "dur": 81.987,
    "args": {
      "External id": 2075,      "device": 2, "stream": 0,      "correlation": 1047, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1047, "pid": 2, "tid": 0, "ts": 7253740936155.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936237.642, "dur": 59.335,
    "args": {
      "External id": 2076,      "device": 2, "stream": 0,      "correlation": 1048, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1048, "pid": 2, "tid": 0, "ts": 7253740936237.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740936296.978, "dur": 76.335,
    "args": {
      "External id": 2077,      "device": 2, "stream": 0,      "correlation": 1049, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1049, "pid": 2, "tid": 0, "ts": 7253740936296.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936373.314, "dur": 63.104,
    "args": {
      "External id": 2078,      "device": 2, "stream": 0,      "correlation": 1050, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1050, "pid": 2, "tid": 0, "ts": 7253740936373.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740936436.419, "dur": 75.492,
    "args": {
      "External id": 2080,      "device": 2, "stream": 0,      "correlation": 1051, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1051, "pid": 2, "tid": 0, "ts": 7253740936436.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740936511.912, "dur": 81.186,
    "args": {
      "External id": 2081,      "device": 2, "stream": 0,      "correlation": 1052, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1052, "pid": 2, "tid": 0, "ts": 7253740936511.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936599.874, "dur": 45.383,
    "args": {
      "External id": 2083,      "device": 2, "stream": 0,      "correlation": 1053, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1053, "pid": 2, "tid": 0, "ts": 7253740936599.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740936645.258, "dur": 77.698,
    "args": {
      "External id": 2086,      "device": 2, "stream": 0,      "correlation": 1054, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1727.4061184828633
    }
  },
  {
    "ph": "f", "id": 1054, "pid": 2, "tid": 0, "ts": 7253740936645.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936729.933, "dur": 5.451,
    "args": {
      "External id": 2093,      "device": 2, "stream": 0,      "correlation": 1055, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1055, "pid": 2, "tid": 0, "ts": 7253740936729.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936735.385, "dur": 5.692,
    "args": {
      "External id": 2094,      "device": 2, "stream": 0,      "correlation": 1056, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1056, "pid": 2, "tid": 0, "ts": 7253740936735.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740936741.078, "dur": 81.867,
    "args": {
      "External id": 2095,      "device": 2, "stream": 0,      "correlation": 1057, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1057, "pid": 2, "tid": 0, "ts": 7253740936741.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936822.946, "dur": 65.550,
    "args": {
      "External id": 2096,      "device": 2, "stream": 0,      "correlation": 1058, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1058, "pid": 2, "tid": 0, "ts": 7253740936822.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740936888.497, "dur": 79.181,
    "args": {
      "External id": 2097,      "device": 2, "stream": 0,      "correlation": 1059, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1059, "pid": 2, "tid": 0, "ts": 7253740936888.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740936967.679, "dur": 63.344,
    "args": {
      "External id": 2098,      "device": 2, "stream": 0,      "correlation": 1060, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1060, "pid": 2, "tid": 0, "ts": 7253740936967.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740937031.024, "dur": 76.014,
    "args": {
      "External id": 2100,      "device": 2, "stream": 0,      "correlation": 1061, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1061, "pid": 2, "tid": 0, "ts": 7253740937031.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740937107.039, "dur": 84.673,
    "args": {
      "External id": 2101,      "device": 2, "stream": 0,      "correlation": 1062, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1062, "pid": 2, "tid": 0, "ts": 7253740937107.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937198.730, "dur": 44.581,
    "args": {
      "External id": 2103,      "device": 2, "stream": 0,      "correlation": 1063, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1063, "pid": 2, "tid": 0, "ts": 7253740937198.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740937243.312, "dur": 78.299,
    "args": {
      "External id": 2106,      "device": 2, "stream": 0,      "correlation": 1064, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1714.1472286079184
    }
  },
  {
    "ph": "f", "id": 1064, "pid": 2, "tid": 0, "ts": 7253740937243.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937328.668, "dur": 5.371,
    "args": {
      "External id": 2113,      "device": 2, "stream": 0,      "correlation": 1065, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1065, "pid": 2, "tid": 0, "ts": 7253740937328.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937334.040, "dur": 5.572,
    "args": {
      "External id": 2114,      "device": 2, "stream": 0,      "correlation": 1066, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1066, "pid": 2, "tid": 0, "ts": 7253740937334.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740937339.613, "dur": 84.152,
    "args": {
      "External id": 2115,      "device": 2, "stream": 0,      "correlation": 1067, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1067, "pid": 2, "tid": 0, "ts": 7253740937339.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937423.766, "dur": 62.142,
    "args": {
      "External id": 2116,      "device": 2, "stream": 0,      "correlation": 1068, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1068, "pid": 2, "tid": 0, "ts": 7253740937423.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740937485.909, "dur": 77.016,
    "args": {
      "External id": 2117,      "device": 2, "stream": 0,      "correlation": 1069, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1069, "pid": 2, "tid": 0, "ts": 7253740937485.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937562.926, "dur": 65.630,
    "args": {
      "External id": 2118,      "device": 2, "stream": 0,      "correlation": 1070, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1070, "pid": 2, "tid": 0, "ts": 7253740937562.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740937628.557, "dur": 76.414,
    "args": {
      "External id": 2120,      "device": 2, "stream": 0,      "correlation": 1071, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1071, "pid": 2, "tid": 0, "ts": 7253740937628.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740937704.972, "dur": 84.353,
    "args": {
      "External id": 2121,      "device": 2, "stream": 0,      "correlation": 1072, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1072, "pid": 2, "tid": 0, "ts": 7253740937704.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937796.382, "dur": 46.346,
    "args": {
      "External id": 2123,      "device": 2, "stream": 0,      "correlation": 1073, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1073, "pid": 2, "tid": 0, "ts": 7253740937796.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740937842.729, "dur": 76.294,
    "args": {
      "External id": 2126,      "device": 2, "stream": 0,      "correlation": 1074, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1759.1942853397995
    }
  },
  {
    "ph": "f", "id": 1074, "pid": 2, "tid": 0, "ts": 7253740937842.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937925.920, "dur": 5.371,
    "args": {
      "External id": 2133,      "device": 2, "stream": 0,      "correlation": 1075, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1075, "pid": 2, "tid": 0, "ts": 7253740937925.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740937931.292, "dur": 5.812,
    "args": {
      "External id": 2134,      "device": 2, "stream": 0,      "correlation": 1076, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1076, "pid": 2, "tid": 0, "ts": 7253740937931.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740937937.105, "dur": 81.948,
    "args": {
      "External id": 2135,      "device": 2, "stream": 0,      "correlation": 1077, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1077, "pid": 2, "tid": 0, "ts": 7253740937937.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938019.054, "dur": 58.092,
    "args": {
      "External id": 2136,      "device": 2, "stream": 0,      "correlation": 1078, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1078, "pid": 2, "tid": 0, "ts": 7253740938019.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740938077.147, "dur": 76.615,
    "args": {
      "External id": 2137,      "device": 2, "stream": 0,      "correlation": 1079, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1079, "pid": 2, "tid": 0, "ts": 7253740938077.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938153.763, "dur": 63.906,
    "args": {
      "External id": 2138,      "device": 2, "stream": 0,      "correlation": 1080, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1080, "pid": 2, "tid": 0, "ts": 7253740938153.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740938217.670, "dur": 75.051,
    "args": {
      "External id": 2140,      "device": 2, "stream": 0,      "correlation": 1081, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1081, "pid": 2, "tid": 0, "ts": 7253740938217.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740938294.117, "dur": 82.308,
    "args": {
      "External id": 2141,      "device": 2, "stream": 0,      "correlation": 1082, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1082, "pid": 2, "tid": 0, "ts": 7253740938294.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938383.362, "dur": 43.620,
    "args": {
      "External id": 2143,      "device": 2, "stream": 0,      "correlation": 1083, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1083, "pid": 2, "tid": 0, "ts": 7253740938383.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740938426.983, "dur": 78.057,
    "args": {
      "External id": 2146,      "device": 2, "stream": 0,      "correlation": 1084, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1719.4615286069334
    }
  },
  {
    "ph": "f", "id": 1084, "pid": 2, "tid": 0, "ts": 7253740938426.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938512.177, "dur": 5.813,
    "args": {
      "External id": 2153,      "device": 2, "stream": 0,      "correlation": 1085, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1085, "pid": 2, "tid": 0, "ts": 7253740938512.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938517.991, "dur": 6.093,
    "args": {
      "External id": 2154,      "device": 2, "stream": 0,      "correlation": 1086, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1086, "pid": 2, "tid": 0, "ts": 7253740938517.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740938524.085, "dur": 81.907,
    "args": {
      "External id": 2155,      "device": 2, "stream": 0,      "correlation": 1087, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1087, "pid": 2, "tid": 0, "ts": 7253740938524.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938605.993, "dur": 62.943,
    "args": {
      "External id": 2156,      "device": 2, "stream": 0,      "correlation": 1088, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1088, "pid": 2, "tid": 0, "ts": 7253740938605.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740938668.937, "dur": 76.615,
    "args": {
      "External id": 2157,      "device": 2, "stream": 0,      "correlation": 1089, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1089, "pid": 2, "tid": 0, "ts": 7253740938668.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938745.553, "dur": 65.149,
    "args": {
      "External id": 2158,      "device": 2, "stream": 0,      "correlation": 1090, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1090, "pid": 2, "tid": 0, "ts": 7253740938745.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740938810.703, "dur": 75.172,
    "args": {
      "External id": 2160,      "device": 2, "stream": 0,      "correlation": 1091, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1091, "pid": 2, "tid": 0, "ts": 7253740938810.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740938885.876, "dur": 81.626,
    "args": {
      "External id": 2161,      "device": 2, "stream": 0,      "correlation": 1092, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1092, "pid": 2, "tid": 0, "ts": 7253740938885.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740938974.439, "dur": 45.143,
    "args": {
      "External id": 2163,      "device": 2, "stream": 0,      "correlation": 1093, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1093, "pid": 2, "tid": 0, "ts": 7253740938974.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740939019.583, "dur": 77.016,
    "args": {
      "External id": 2166,      "device": 2, "stream": 0,      "correlation": 1094, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1742.7026240959788
    }
  },
  {
    "ph": "f", "id": 1094, "pid": 2, "tid": 0, "ts": 7253740939019.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939103.656, "dur": 5.291,
    "args": {
      "External id": 2173,      "device": 2, "stream": 0,      "correlation": 1095, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1095, "pid": 2, "tid": 0, "ts": 7253740939103.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939108.948, "dur": 5.732,
    "args": {
      "External id": 2174,      "device": 2, "stream": 0,      "correlation": 1096, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1096, "pid": 2, "tid": 0, "ts": 7253740939108.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740939114.681, "dur": 81.827,
    "args": {
      "External id": 2175,      "device": 2, "stream": 0,      "correlation": 1097, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1097, "pid": 2, "tid": 0, "ts": 7253740939114.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939196.509, "dur": 61.340,
    "args": {
      "External id": 2176,      "device": 2, "stream": 0,      "correlation": 1098, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1098, "pid": 2, "tid": 0, "ts": 7253740939196.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740939257.850, "dur": 77.617,
    "args": {
      "External id": 2177,      "device": 2, "stream": 0,      "correlation": 1099, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1099, "pid": 2, "tid": 0, "ts": 7253740939257.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939335.468, "dur": 63.465,
    "args": {
      "External id": 2178,      "device": 2, "stream": 0,      "correlation": 1100, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1100, "pid": 2, "tid": 0, "ts": 7253740939335.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740939398.934, "dur": 74.530,
    "args": {
      "External id": 2180,      "device": 2, "stream": 0,      "correlation": 1101, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1101, "pid": 2, "tid": 0, "ts": 7253740939398.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740939473.465, "dur": 81.667,
    "args": {
      "External id": 2181,      "device": 2, "stream": 0,      "correlation": 1102, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1102, "pid": 2, "tid": 0, "ts": 7253740939473.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939562.389, "dur": 45.143,
    "args": {
      "External id": 2183,      "device": 2, "stream": 0,      "correlation": 1103, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1103, "pid": 2, "tid": 0, "ts": 7253740939562.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740939607.533, "dur": 79.863,
    "args": {
      "External id": 2186,      "device": 2, "stream": 0,      "correlation": 1104, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1680.5785835921065
    }
  },
  {
    "ph": "f", "id": 1104, "pid": 2, "tid": 0, "ts": 7253740939607.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939694.252, "dur": 5.412,
    "args": {
      "External id": 2193,      "device": 2, "stream": 0,      "correlation": 1105, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1105, "pid": 2, "tid": 0, "ts": 7253740939694.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939699.665, "dur": 5.251,
    "args": {
      "External id": 2194,      "device": 2, "stream": 0,      "correlation": 1106, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1106, "pid": 2, "tid": 0, "ts": 7253740939699.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740939704.917, "dur": 81.787,
    "args": {
      "External id": 2195,      "device": 2, "stream": 0,      "correlation": 1107, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1107, "pid": 2, "tid": 0, "ts": 7253740939704.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939786.705, "dur": 65.269,
    "args": {
      "External id": 2196,      "device": 2, "stream": 0,      "correlation": 1108, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1108, "pid": 2, "tid": 0, "ts": 7253740939786.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740939851.975, "dur": 77.056,
    "args": {
      "External id": 2197,      "device": 2, "stream": 0,      "correlation": 1109, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1109, "pid": 2, "tid": 0, "ts": 7253740939851.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740939929.032, "dur": 64.066,
    "args": {
      "External id": 2198,      "device": 2, "stream": 0,      "correlation": 1110, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1110, "pid": 2, "tid": 0, "ts": 7253740939929.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740939993.099, "dur": 75.252,
    "args": {
      "External id": 2200,      "device": 2, "stream": 0,      "correlation": 1111, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1111, "pid": 2, "tid": 0, "ts": 7253740939993.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740940068.352, "dur": 82.668,
    "args": {
      "External id": 2201,      "device": 2, "stream": 0,      "correlation": 1112, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1112, "pid": 2, "tid": 0, "ts": 7253740940068.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940157.797, "dur": 44.501,
    "args": {
      "External id": 2203,      "device": 2, "stream": 0,      "correlation": 1113, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1113, "pid": 2, "tid": 0, "ts": 7253740940157.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740940202.299, "dur": 78.540,
    "args": {
      "External id": 2206,      "device": 2, "stream": 0,      "correlation": 1114, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1708.8874345883041
    }
  },
  {
    "ph": "f", "id": 1114, "pid": 2, "tid": 0, "ts": 7253740940202.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940287.896, "dur": 5.331,
    "args": {
      "External id": 2213,      "device": 2, "stream": 0,      "correlation": 1115, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1115, "pid": 2, "tid": 0, "ts": 7253740940287.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940293.228, "dur": 6.133,
    "args": {
      "External id": 2214,      "device": 2, "stream": 0,      "correlation": 1116, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1116, "pid": 2, "tid": 0, "ts": 7253740940293.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740940299.362, "dur": 82.188,
    "args": {
      "External id": 2215,      "device": 2, "stream": 0,      "correlation": 1117, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1117, "pid": 2, "tid": 0, "ts": 7253740940299.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940381.551, "dur": 62.543,
    "args": {
      "External id": 2216,      "device": 2, "stream": 0,      "correlation": 1118, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1118, "pid": 2, "tid": 0, "ts": 7253740940381.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740940444.095, "dur": 77.537,
    "args": {
      "External id": 2217,      "device": 2, "stream": 0,      "correlation": 1119, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1119, "pid": 2, "tid": 0, "ts": 7253740940444.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940521.633, "dur": 64.788,
    "args": {
      "External id": 2218,      "device": 2, "stream": 0,      "correlation": 1120, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1120, "pid": 2, "tid": 0, "ts": 7253740940521.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740940586.422, "dur": 74.891,
    "args": {
      "External id": 2220,      "device": 2, "stream": 0,      "correlation": 1121, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1121, "pid": 2, "tid": 0, "ts": 7253740940586.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740940661.314, "dur": 81.706,
    "args": {
      "External id": 2221,      "device": 2, "stream": 0,      "correlation": 1122, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1122, "pid": 2, "tid": 0, "ts": 7253740940661.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940749.797, "dur": 44.341,
    "args": {
      "External id": 2223,      "device": 2, "stream": 0,      "correlation": 1123, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1123, "pid": 2, "tid": 0, "ts": 7253740940749.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740940794.139, "dur": 77.497,
    "args": {
      "External id": 2226,      "device": 2, "stream": 0,      "correlation": 1124, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1731.8863454540763
    }
  },
  {
    "ph": "f", "id": 1124, "pid": 2, "tid": 0, "ts": 7253740940794.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940878.893, "dur": 5.251,
    "args": {
      "External id": 2233,      "device": 2, "stream": 0,      "correlation": 1125, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1125, "pid": 2, "tid": 0, "ts": 7253740940878.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940884.145, "dur": 5.813,
    "args": {
      "External id": 2234,      "device": 2, "stream": 0,      "correlation": 1126, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1126, "pid": 2, "tid": 0, "ts": 7253740940884.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740940889.959, "dur": 81.265,
    "args": {
      "External id": 2235,      "device": 2, "stream": 0,      "correlation": 1127, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1127, "pid": 2, "tid": 0, "ts": 7253740940889.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740940971.225, "dur": 59.095,
    "args": {
      "External id": 2236,      "device": 2, "stream": 0,      "correlation": 1128, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1128, "pid": 2, "tid": 0, "ts": 7253740940971.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740941030.321, "dur": 75.573,
    "args": {
      "External id": 2237,      "device": 2, "stream": 0,      "correlation": 1129, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1129, "pid": 2, "tid": 0, "ts": 7253740941030.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941105.895, "dur": 64.066,
    "args": {
      "External id": 2238,      "device": 2, "stream": 0,      "correlation": 1130, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1130, "pid": 2, "tid": 0, "ts": 7253740941105.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740941169.962, "dur": 74.851,
    "args": {
      "External id": 2240,      "device": 2, "stream": 0,      "correlation": 1131, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1131, "pid": 2, "tid": 0, "ts": 7253740941169.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740941244.814, "dur": 80.263,
    "args": {
      "External id": 2241,      "device": 2, "stream": 0,      "correlation": 1132, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1132, "pid": 2, "tid": 0, "ts": 7253740941244.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941332.134, "dur": 45.464,
    "args": {
      "External id": 2243,      "device": 2, "stream": 0,      "correlation": 1133, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1133, "pid": 2, "tid": 0, "ts": 7253740941332.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740941377.599, "dur": 78.018,
    "args": {
      "External id": 2246,      "device": 2, "stream": 0,      "correlation": 1134, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1720.321050000641
    }
  },
  {
    "ph": "f", "id": 1134, "pid": 2, "tid": 0, "ts": 7253740941377.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941463.516, "dur": 5.211,
    "args": {
      "External id": 2253,      "device": 2, "stream": 0,      "correlation": 1135, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1135, "pid": 2, "tid": 0, "ts": 7253740941463.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941468.728, "dur": 5.652,
    "args": {
      "External id": 2254,      "device": 2, "stream": 0,      "correlation": 1136, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1136, "pid": 2, "tid": 0, "ts": 7253740941468.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740941474.381, "dur": 80.704,
    "args": {
      "External id": 2255,      "device": 2, "stream": 0,      "correlation": 1137, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1137, "pid": 2, "tid": 0, "ts": 7253740941474.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941555.086, "dur": 62.303,
    "args": {
      "External id": 2256,      "device": 2, "stream": 0,      "correlation": 1138, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1138, "pid": 2, "tid": 0, "ts": 7253740941555.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740941617.390, "dur": 76.334,
    "args": {
      "External id": 2257,      "device": 2, "stream": 0,      "correlation": 1139, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1139, "pid": 2, "tid": 0, "ts": 7253740941617.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941693.725, "dur": 62.382,
    "args": {
      "External id": 2258,      "device": 2, "stream": 0,      "correlation": 1140, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1140, "pid": 2, "tid": 0, "ts": 7253740941693.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740941756.108, "dur": 74.811,
    "args": {
      "External id": 2260,      "device": 2, "stream": 0,      "correlation": 1141, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1141, "pid": 2, "tid": 0, "ts": 7253740941756.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740941830.920, "dur": 80.825,
    "args": {
      "External id": 2261,      "device": 2, "stream": 0,      "correlation": 1142, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1142, "pid": 2, "tid": 0, "ts": 7253740941830.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740941918.641, "dur": 45.825,
    "args": {
      "External id": 2263,      "device": 2, "stream": 0,      "correlation": 1143, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1143, "pid": 2, "tid": 0, "ts": 7253740941918.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740941964.467, "dur": 76.895,
    "args": {
      "External id": 2266,      "device": 2, "stream": 0,      "correlation": 1144, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1745.444860590928
    }
  },
  {
    "ph": "f", "id": 1144, "pid": 2, "tid": 0, "ts": 7253740941964.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942048.179, "dur": 5.331,
    "args": {
      "External id": 2273,      "device": 2, "stream": 0,      "correlation": 1145, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1145, "pid": 2, "tid": 0, "ts": 7253740942048.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942053.511, "dur": 5.372,
    "args": {
      "External id": 2274,      "device": 2, "stream": 0,      "correlation": 1146, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1146, "pid": 2, "tid": 0, "ts": 7253740942053.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740942058.884, "dur": 81.666,
    "args": {
      "External id": 2275,      "device": 2, "stream": 0,      "correlation": 1147, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1147, "pid": 2, "tid": 0, "ts": 7253740942058.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942140.551, "dur": 65.069,
    "args": {
      "External id": 2276,      "device": 2, "stream": 0,      "correlation": 1148, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1148, "pid": 2, "tid": 0, "ts": 7253740942140.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740942205.621, "dur": 76.294,
    "args": {
      "External id": 2277,      "device": 2, "stream": 0,      "correlation": 1149, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1149, "pid": 2, "tid": 0, "ts": 7253740942205.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942281.916, "dur": 64.627,
    "args": {
      "External id": 2278,      "device": 2, "stream": 0,      "correlation": 1150, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1150, "pid": 2, "tid": 0, "ts": 7253740942281.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740942346.544, "dur": 74.130,
    "args": {
      "External id": 2280,      "device": 2, "stream": 0,      "correlation": 1151, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1151, "pid": 2, "tid": 0, "ts": 7253740942346.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740942420.675, "dur": 81.827,
    "args": {
      "External id": 2281,      "device": 2, "stream": 0,      "correlation": 1152, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1152, "pid": 2, "tid": 0, "ts": 7253740942420.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942509.438, "dur": 45.263,
    "args": {
      "External id": 2283,      "device": 2, "stream": 0,      "correlation": 1153, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1153, "pid": 2, "tid": 0, "ts": 7253740942509.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740942554.702, "dur": 77.297,
    "args": {
      "External id": 2286,      "device": 2, "stream": 0,      "correlation": 1154, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1736.3674092473286
    }
  },
  {
    "ph": "f", "id": 1154, "pid": 2, "tid": 0, "ts": 7253740942554.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942639.257, "dur": 5.251,
    "args": {
      "External id": 2293,      "device": 2, "stream": 0,      "correlation": 1155, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1155, "pid": 2, "tid": 0, "ts": 7253740942639.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942644.509, "dur": 5.772,
    "args": {
      "External id": 2294,      "device": 2, "stream": 0,      "correlation": 1156, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1156, "pid": 2, "tid": 0, "ts": 7253740942644.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740942650.282, "dur": 81.747,
    "args": {
      "External id": 2295,      "device": 2, "stream": 0,      "correlation": 1157, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1157, "pid": 2, "tid": 0, "ts": 7253740942650.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942732.030, "dur": 61.741,
    "args": {
      "External id": 2296,      "device": 2, "stream": 0,      "correlation": 1158, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1158, "pid": 2, "tid": 0, "ts": 7253740942732.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740942793.772, "dur": 77.737,
    "args": {
      "External id": 2297,      "device": 2, "stream": 0,      "correlation": 1159, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1159, "pid": 2, "tid": 0, "ts": 7253740942793.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740942871.510, "dur": 64.547,
    "args": {
      "External id": 2298,      "device": 2, "stream": 0,      "correlation": 1160, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1160, "pid": 2, "tid": 0, "ts": 7253740942871.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740942936.058, "dur": 74.170,
    "args": {
      "External id": 2300,      "device": 2, "stream": 0,      "correlation": 1161, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1161, "pid": 2, "tid": 0, "ts": 7253740942936.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740943010.229, "dur": 80.664,
    "args": {
      "External id": 2301,      "device": 2, "stream": 0,      "correlation": 1162, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1162, "pid": 2, "tid": 0, "ts": 7253740943010.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943098.391, "dur": 45.624,
    "args": {
      "External id": 2303,      "device": 2, "stream": 0,      "correlation": 1163, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1163, "pid": 2, "tid": 0, "ts": 7253740943098.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740943144.016, "dur": 77.176,
    "args": {
      "External id": 2306,      "device": 2, "stream": 0,      "correlation": 1164, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1739.089728805214
    }
  },
  {
    "ph": "f", "id": 1164, "pid": 2, "tid": 0, "ts": 7253740943144.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943228.290, "dur": 5.291,
    "args": {
      "External id": 2313,      "device": 2, "stream": 0,      "correlation": 1165, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1165, "pid": 2, "tid": 0, "ts": 7253740943228.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943233.582, "dur": 5.612,
    "args": {
      "External id": 2314,      "device": 2, "stream": 0,      "correlation": 1166, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1166, "pid": 2, "tid": 0, "ts": 7253740943233.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740943239.195, "dur": 80.343,
    "args": {
      "External id": 2315,      "device": 2, "stream": 0,      "correlation": 1167, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1167, "pid": 2, "tid": 0, "ts": 7253740943239.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943319.539, "dur": 60.498,
    "args": {
      "External id": 2316,      "device": 2, "stream": 0,      "correlation": 1168, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1168, "pid": 2, "tid": 0, "ts": 7253740943319.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740943380.038, "dur": 76.495,
    "args": {
      "External id": 2317,      "device": 2, "stream": 0,      "correlation": 1169, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1169, "pid": 2, "tid": 0, "ts": 7253740943380.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943456.534, "dur": 63.224,
    "args": {
      "External id": 2318,      "device": 2, "stream": 0,      "correlation": 1170, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1170, "pid": 2, "tid": 0, "ts": 7253740943456.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740943519.759, "dur": 74.250,
    "args": {
      "External id": 2320,      "device": 2, "stream": 0,      "correlation": 1171, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1171, "pid": 2, "tid": 0, "ts": 7253740943519.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740943594.010, "dur": 80.383,
    "args": {
      "External id": 2321,      "device": 2, "stream": 0,      "correlation": 1172, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1172, "pid": 2, "tid": 0, "ts": 7253740943594.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943682.052, "dur": 45.263,
    "args": {
      "External id": 2323,      "device": 2, "stream": 0,      "correlation": 1173, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1173, "pid": 2, "tid": 0, "ts": 7253740943682.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740943727.316, "dur": 79.742,
    "args": {
      "External id": 2326,      "device": 2, "stream": 0,      "correlation": 1174, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1683.128650790665
    }
  },
  {
    "ph": "f", "id": 1174, "pid": 2, "tid": 0, "ts": 7253740943727.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943814.195, "dur": 5.291,
    "args": {
      "External id": 2333,      "device": 2, "stream": 0,      "correlation": 1175, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1175, "pid": 2, "tid": 0, "ts": 7253740943814.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943819.487, "dur": 5.652,
    "args": {
      "External id": 2334,      "device": 2, "stream": 0,      "correlation": 1176, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1176, "pid": 2, "tid": 0, "ts": 7253740943819.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740943825.140, "dur": 80.344,
    "args": {
      "External id": 2335,      "device": 2, "stream": 0,      "correlation": 1177, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1177, "pid": 2, "tid": 0, "ts": 7253740943825.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740943905.485, "dur": 58.654,
    "args": {
      "External id": 2336,      "device": 2, "stream": 0,      "correlation": 1178, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1178, "pid": 2, "tid": 0, "ts": 7253740943905.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740943964.140, "dur": 76.775,
    "args": {
      "External id": 2337,      "device": 2, "stream": 0,      "correlation": 1179, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1179, "pid": 2, "tid": 0, "ts": 7253740943964.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944040.916, "dur": 62.663,
    "args": {
      "External id": 2338,      "device": 2, "stream": 0,      "correlation": 1180, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1180, "pid": 2, "tid": 0, "ts": 7253740944040.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740944103.580, "dur": 75.011,
    "args": {
      "External id": 2340,      "device": 2, "stream": 0,      "correlation": 1181, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1181, "pid": 2, "tid": 0, "ts": 7253740944103.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740944178.592, "dur": 80.504,
    "args": {
      "External id": 2341,      "device": 2, "stream": 0,      "correlation": 1182, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1182, "pid": 2, "tid": 0, "ts": 7253740944178.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944266.073, "dur": 45.424,
    "args": {
      "External id": 2343,      "device": 2, "stream": 0,      "correlation": 1183, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1183, "pid": 2, "tid": 0, "ts": 7253740944266.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740944311.498, "dur": 78.018,
    "args": {
      "External id": 2346,      "device": 2, "stream": 0,      "correlation": 1184, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1720.321050000641
    }
  },
  {
    "ph": "f", "id": 1184, "pid": 2, "tid": 0, "ts": 7253740944311.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944396.373, "dur": 5.291,
    "args": {
      "External id": 2353,      "device": 2, "stream": 0,      "correlation": 1185, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1185, "pid": 2, "tid": 0, "ts": 7253740944396.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944401.665, "dur": 6.133,
    "args": {
      "External id": 2354,      "device": 2, "stream": 0,      "correlation": 1186, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1186, "pid": 2, "tid": 0, "ts": 7253740944401.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740944407.799, "dur": 80.343,
    "args": {
      "External id": 2355,      "device": 2, "stream": 0,      "correlation": 1187, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1187, "pid": 2, "tid": 0, "ts": 7253740944407.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944488.143, "dur": 62.623,
    "args": {
      "External id": 2356,      "device": 2, "stream": 0,      "correlation": 1188, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1188, "pid": 2, "tid": 0, "ts": 7253740944488.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740944550.767, "dur": 76.174,
    "args": {
      "External id": 2357,      "device": 2, "stream": 0,      "correlation": 1189, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1189, "pid": 2, "tid": 0, "ts": 7253740944550.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944626.942, "dur": 65.991,
    "args": {
      "External id": 2358,      "device": 2, "stream": 0,      "correlation": 1190, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1190, "pid": 2, "tid": 0, "ts": 7253740944626.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740944692.934, "dur": 74.169,
    "args": {
      "External id": 2360,      "device": 2, "stream": 0,      "correlation": 1191, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1191, "pid": 2, "tid": 0, "ts": 7253740944692.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740944767.104, "dur": 80.223,
    "args": {
      "External id": 2361,      "device": 2, "stream": 0,      "correlation": 1192, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1192, "pid": 2, "tid": 0, "ts": 7253740944767.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944854.425, "dur": 45.102,
    "args": {
      "External id": 2363,      "device": 2, "stream": 0,      "correlation": 1193, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1193, "pid": 2, "tid": 0, "ts": 7253740944854.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740944899.528, "dur": 77.577,
    "args": {
      "External id": 2366,      "device": 2, "stream": 0,      "correlation": 1194, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1730.1003892856222
    }
  },
  {
    "ph": "f", "id": 1194, "pid": 2, "tid": 0, "ts": 7253740944899.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944984.564, "dur": 5.291,
    "args": {
      "External id": 2373,      "device": 2, "stream": 0,      "correlation": 1195, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1195, "pid": 2, "tid": 0, "ts": 7253740944984.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740944989.856, "dur": 5.732,
    "args": {
      "External id": 2374,      "device": 2, "stream": 0,      "correlation": 1196, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1196, "pid": 2, "tid": 0, "ts": 7253740944989.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740944995.589, "dur": 81.707,
    "args": {
      "External id": 2375,      "device": 2, "stream": 0,      "correlation": 1197, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1197, "pid": 2, "tid": 0, "ts": 7253740944995.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945077.297, "dur": 61.780,
    "args": {
      "External id": 2376,      "device": 2, "stream": 0,      "correlation": 1198, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1198, "pid": 2, "tid": 0, "ts": 7253740945077.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740945139.078, "dur": 77.457,
    "args": {
      "External id": 2377,      "device": 2, "stream": 0,      "correlation": 1199, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1199, "pid": 2, "tid": 0, "ts": 7253740945139.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945216.536, "dur": 64.187,
    "args": {
      "External id": 2378,      "device": 2, "stream": 0,      "correlation": 1200, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1200, "pid": 2, "tid": 0, "ts": 7253740945216.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740945280.724, "dur": 75.492,
    "args": {
      "External id": 2380,      "device": 2, "stream": 0,      "correlation": 1201, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1201, "pid": 2, "tid": 0, "ts": 7253740945280.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740945356.217, "dur": 81.827,
    "args": {
      "External id": 2381,      "device": 2, "stream": 0,      "correlation": 1202, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1202, "pid": 2, "tid": 0, "ts": 7253740945356.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945445.141, "dur": 45.223,
    "args": {
      "External id": 2383,      "device": 2, "stream": 0,      "correlation": 1203, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1203, "pid": 2, "tid": 0, "ts": 7253740945445.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740945490.365, "dur": 78.860,
    "args": {
      "External id": 2386,      "device": 2, "stream": 0,      "correlation": 1204, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1701.953158088282
    }
  },
  {
    "ph": "f", "id": 1204, "pid": 2, "tid": 0, "ts": 7253740945490.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945576.243, "dur": 5.371,
    "args": {
      "External id": 2393,      "device": 2, "stream": 0,      "correlation": 1205, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1205, "pid": 2, "tid": 0, "ts": 7253740945576.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945581.615, "dur": 5.772,
    "args": {
      "External id": 2394,      "device": 2, "stream": 0,      "correlation": 1206, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1206, "pid": 2, "tid": 0, "ts": 7253740945581.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740945587.388, "dur": 81.627,
    "args": {
      "External id": 2395,      "device": 2, "stream": 0,      "correlation": 1207, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1207, "pid": 2, "tid": 0, "ts": 7253740945587.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945669.016, "dur": 63.424,
    "args": {
      "External id": 2396,      "device": 2, "stream": 0,      "correlation": 1208, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1208, "pid": 2, "tid": 0, "ts": 7253740945669.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740945732.441, "dur": 78.460,
    "args": {
      "External id": 2397,      "device": 2, "stream": 0,      "correlation": 1209, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1209, "pid": 2, "tid": 0, "ts": 7253740945732.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740945810.902, "dur": 65.108,
    "args": {
      "External id": 2398,      "device": 2, "stream": 0,      "correlation": 1210, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1210, "pid": 2, "tid": 0, "ts": 7253740945810.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740945876.011, "dur": 75.452,
    "args": {
      "External id": 2400,      "device": 2, "stream": 0,      "correlation": 1211, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1211, "pid": 2, "tid": 0, "ts": 7253740945876.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740945951.464, "dur": 81.627,
    "args": {
      "External id": 2401,      "device": 2, "stream": 0,      "correlation": 1212, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1212, "pid": 2, "tid": 0, "ts": 7253740945951.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946040.188, "dur": 44.702,
    "args": {
      "External id": 2403,      "device": 2, "stream": 0,      "correlation": 1213, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1213, "pid": 2, "tid": 0, "ts": 7253740946040.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740946084.891, "dur": 74.650,
    "args": {
      "External id": 2406,      "device": 2, "stream": 0,      "correlation": 1214, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1797.9361026643983
    }
  },
  {
    "ph": "f", "id": 1214, "pid": 2, "tid": 0, "ts": 7253740946084.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946166.398, "dur": 5.812,
    "args": {
      "External id": 2413,      "device": 2, "stream": 0,      "correlation": 1215, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1215, "pid": 2, "tid": 0, "ts": 7253740946166.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946172.211, "dur": 6.815,
    "args": {
      "External id": 2414,      "device": 2, "stream": 0,      "correlation": 1216, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1216, "pid": 2, "tid": 0, "ts": 7253740946172.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740946179.027, "dur": 82.469,
    "args": {
      "External id": 2415,      "device": 2, "stream": 0,      "correlation": 1217, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1217, "pid": 2, "tid": 0, "ts": 7253740946179.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946261.497, "dur": 62.743,
    "args": {
      "External id": 2416,      "device": 2, "stream": 0,      "correlation": 1218, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1218, "pid": 2, "tid": 0, "ts": 7253740946261.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740946324.241, "dur": 78.499,
    "args": {
      "External id": 2417,      "device": 2, "stream": 0,      "correlation": 1219, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1219, "pid": 2, "tid": 0, "ts": 7253740946324.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946402.741, "dur": 62.663,
    "args": {
      "External id": 2418,      "device": 2, "stream": 0,      "correlation": 1220, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1220, "pid": 2, "tid": 0, "ts": 7253740946402.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740946465.405, "dur": 74.851,
    "args": {
      "External id": 2420,      "device": 2, "stream": 0,      "correlation": 1221, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1221, "pid": 2, "tid": 0, "ts": 7253740946465.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740946540.257, "dur": 81.947,
    "args": {
      "External id": 2421,      "device": 2, "stream": 0,      "correlation": 1222, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1222, "pid": 2, "tid": 0, "ts": 7253740946540.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946629.782, "dur": 45.624,
    "args": {
      "External id": 2423,      "device": 2, "stream": 0,      "correlation": 1223, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1223, "pid": 2, "tid": 0, "ts": 7253740946629.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740946675.407, "dur": 79.301,
    "args": {
      "External id": 2426,      "device": 2, "stream": 0,      "correlation": 1224, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1692.4885627096417
    }
  },
  {
    "ph": "f", "id": 1224, "pid": 2, "tid": 0, "ts": 7253740946675.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946762.808, "dur": 5.331,
    "args": {
      "External id": 2433,      "device": 2, "stream": 0,      "correlation": 1225, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1225, "pid": 2, "tid": 0, "ts": 7253740946762.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946768.140, "dur": 5.492,
    "args": {
      "External id": 2434,      "device": 2, "stream": 0,      "correlation": 1226, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1226, "pid": 2, "tid": 0, "ts": 7253740946768.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740946773.633, "dur": 82.228,
    "args": {
      "External id": 2435,      "device": 2, "stream": 0,      "correlation": 1227, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1227, "pid": 2, "tid": 0, "ts": 7253740946773.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946855.862, "dur": 60.337,
    "args": {
      "External id": 2436,      "device": 2, "stream": 0,      "correlation": 1228, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1228, "pid": 2, "tid": 0, "ts": 7253740946855.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740946916.200, "dur": 75.813,
    "args": {
      "External id": 2437,      "device": 2, "stream": 0,      "correlation": 1229, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1229, "pid": 2, "tid": 0, "ts": 7253740946916.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740946992.014, "dur": 61.501,
    "args": {
      "External id": 2438,      "device": 2, "stream": 0,      "correlation": 1230, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1230, "pid": 2, "tid": 0, "ts": 7253740946992.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740947053.516, "dur": 74.209,
    "args": {
      "External id": 2440,      "device": 2, "stream": 0,      "correlation": 1231, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1231, "pid": 2, "tid": 0, "ts": 7253740947053.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740947127.726, "dur": 80.745,
    "args": {
      "External id": 2441,      "device": 2, "stream": 0,      "correlation": 1232, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1232, "pid": 2, "tid": 0, "ts": 7253740947127.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947215.528, "dur": 45.183,
    "args": {
      "External id": 2443,      "device": 2, "stream": 0,      "correlation": 1233, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1233, "pid": 2, "tid": 0, "ts": 7253740947215.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740947260.712, "dur": 77.497,
    "args": {
      "External id": 2446,      "device": 2, "stream": 0,      "correlation": 1234, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1731.8863454540763
    }
  },
  {
    "ph": "f", "id": 1234, "pid": 2, "tid": 0, "ts": 7253740947260.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947345.146, "dur": 5.251,
    "args": {
      "External id": 2453,      "device": 2, "stream": 0,      "correlation": 1235, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1235, "pid": 2, "tid": 0, "ts": 7253740947345.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947350.398, "dur": 5.772,
    "args": {
      "External id": 2454,      "device": 2, "stream": 0,      "correlation": 1236, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1236, "pid": 2, "tid": 0, "ts": 7253740947350.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740947356.171, "dur": 80.584,
    "args": {
      "External id": 2455,      "device": 2, "stream": 0,      "correlation": 1237, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1237, "pid": 2, "tid": 0, "ts": 7253740947356.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947436.756, "dur": 63.785,
    "args": {
      "External id": 2456,      "device": 2, "stream": 0,      "correlation": 1238, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1238, "pid": 2, "tid": 0, "ts": 7253740947436.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740947500.542, "dur": 75.373,
    "args": {
      "External id": 2457,      "device": 2, "stream": 0,      "correlation": 1239, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1239, "pid": 2, "tid": 0, "ts": 7253740947500.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947575.916, "dur": 62.021,
    "args": {
      "External id": 2458,      "device": 2, "stream": 0,      "correlation": 1240, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1240, "pid": 2, "tid": 0, "ts": 7253740947575.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740947637.938, "dur": 74.290,
    "args": {
      "External id": 2460,      "device": 2, "stream": 0,      "correlation": 1241, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1241, "pid": 2, "tid": 0, "ts": 7253740947637.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740947712.229, "dur": 79.822,
    "args": {
      "External id": 2461,      "device": 2, "stream": 0,      "correlation": 1242, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1242, "pid": 2, "tid": 0, "ts": 7253740947712.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947798.868, "dur": 45.583,
    "args": {
      "External id": 2463,      "device": 2, "stream": 0,      "correlation": 1243, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1243, "pid": 2, "tid": 0, "ts": 7253740947798.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740947844.452, "dur": 77.537,
    "args": {
      "External id": 2466,      "device": 2, "stream": 0,      "correlation": 1244, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1730.9929067038097
    }
  },
  {
    "ph": "f", "id": 1244, "pid": 2, "tid": 0, "ts": 7253740947844.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947928.846, "dur": 5.492,
    "args": {
      "External id": 2473,      "device": 2, "stream": 0,      "correlation": 1245, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1245, "pid": 2, "tid": 0, "ts": 7253740947928.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740947934.339, "dur": 5.572,
    "args": {
      "External id": 2474,      "device": 2, "stream": 0,      "correlation": 1246, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1246, "pid": 2, "tid": 0, "ts": 7253740947934.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740947939.912, "dur": 81.666,
    "args": {
      "External id": 2475,      "device": 2, "stream": 0,      "correlation": 1247, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1247, "pid": 2, "tid": 0, "ts": 7253740947939.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948021.579, "dur": 61.340,
    "args": {
      "External id": 2476,      "device": 2, "stream": 0,      "correlation": 1248, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1248, "pid": 2, "tid": 0, "ts": 7253740948021.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740948082.920, "dur": 76.495,
    "args": {
      "External id": 2477,      "device": 2, "stream": 0,      "correlation": 1249, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1249, "pid": 2, "tid": 0, "ts": 7253740948082.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948159.416, "dur": 61.741,
    "args": {
      "External id": 2478,      "device": 2, "stream": 0,      "correlation": 1250, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1250, "pid": 2, "tid": 0, "ts": 7253740948159.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740948221.158, "dur": 74.570,
    "args": {
      "External id": 2480,      "device": 2, "stream": 0,      "correlation": 1251, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1251, "pid": 2, "tid": 0, "ts": 7253740948221.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740948295.729, "dur": 81.827,
    "args": {
      "External id": 2481,      "device": 2, "stream": 0,      "correlation": 1252, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1252, "pid": 2, "tid": 0, "ts": 7253740948295.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948385.014, "dur": 45.584,
    "args": {
      "External id": 2483,      "device": 2, "stream": 0,      "correlation": 1253, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1253, "pid": 2, "tid": 0, "ts": 7253740948385.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740948430.599, "dur": 77.336,
    "args": {
      "External id": 2486,      "device": 2, "stream": 0,      "correlation": 1254, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1735.4917827171987
    }
  },
  {
    "ph": "f", "id": 1254, "pid": 2, "tid": 0, "ts": 7253740948430.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948514.952, "dur": 5.292,
    "args": {
      "External id": 2493,      "device": 2, "stream": 0,      "correlation": 1255, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1255, "pid": 2, "tid": 0, "ts": 7253740948514.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948520.245, "dur": 6.012,
    "args": {
      "External id": 2494,      "device": 2, "stream": 0,      "correlation": 1256, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1256, "pid": 2, "tid": 0, "ts": 7253740948520.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740948526.258, "dur": 82.389,
    "args": {
      "External id": 2495,      "device": 2, "stream": 0,      "correlation": 1257, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1257, "pid": 2, "tid": 0, "ts": 7253740948526.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948608.648, "dur": 66.712,
    "args": {
      "External id": 2496,      "device": 2, "stream": 0,      "correlation": 1258, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1258, "pid": 2, "tid": 0, "ts": 7253740948608.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740948675.361, "dur": 78.579,
    "args": {
      "External id": 2497,      "device": 2, "stream": 0,      "correlation": 1259, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1259, "pid": 2, "tid": 0, "ts": 7253740948675.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948753.941, "dur": 66.873,
    "args": {
      "External id": 2498,      "device": 2, "stream": 0,      "correlation": 1260, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1260, "pid": 2, "tid": 0, "ts": 7253740948753.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740948820.815, "dur": 75.653,
    "args": {
      "External id": 2500,      "device": 2, "stream": 0,      "correlation": 1261, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1261, "pid": 2, "tid": 0, "ts": 7253740948820.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740948896.469, "dur": 83.350,
    "args": {
      "External id": 2501,      "device": 2, "stream": 0,      "correlation": 1262, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1262, "pid": 2, "tid": 0, "ts": 7253740948896.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740948986.756, "dur": 44.782,
    "args": {
      "External id": 2503,      "device": 2, "stream": 0,      "correlation": 1263, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1263, "pid": 2, "tid": 0, "ts": 7253740948986.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740949031.539, "dur": 78.339,
    "args": {
      "External id": 2506,      "device": 2, "stream": 0,      "correlation": 1264, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1713.271993872862
    }
  },
  {
    "ph": "f", "id": 1264, "pid": 2, "tid": 0, "ts": 7253740949031.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949117.697, "dur": 5.572,
    "args": {
      "External id": 2513,      "device": 2, "stream": 0,      "correlation": 1265, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1265, "pid": 2, "tid": 0, "ts": 7253740949117.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949123.270, "dur": 6.053,
    "args": {
      "External id": 2514,      "device": 2, "stream": 0,      "correlation": 1266, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1266, "pid": 2, "tid": 0, "ts": 7253740949123.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740949129.324, "dur": 83.871,
    "args": {
      "External id": 2515,      "device": 2, "stream": 0,      "correlation": 1267, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1267, "pid": 2, "tid": 0, "ts": 7253740949129.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949213.196, "dur": 64.347,
    "args": {
      "External id": 2516,      "device": 2, "stream": 0,      "correlation": 1268, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1268, "pid": 2, "tid": 0, "ts": 7253740949213.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740949277.544, "dur": 78.660,
    "args": {
      "External id": 2517,      "device": 2, "stream": 0,      "correlation": 1269, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1269, "pid": 2, "tid": 0, "ts": 7253740949277.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949356.205, "dur": 64.547,
    "args": {
      "External id": 2518,      "device": 2, "stream": 0,      "correlation": 1270, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1270, "pid": 2, "tid": 0, "ts": 7253740949356.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740949420.753, "dur": 76.615,
    "args": {
      "External id": 2520,      "device": 2, "stream": 0,      "correlation": 1271, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1271, "pid": 2, "tid": 0, "ts": 7253740949420.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740949497.369, "dur": 83.992,
    "args": {
      "External id": 2521,      "device": 2, "stream": 0,      "correlation": 1272, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1272, "pid": 2, "tid": 0, "ts": 7253740949497.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949588.498, "dur": 44.742,
    "args": {
      "External id": 2523,      "device": 2, "stream": 0,      "correlation": 1273, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1273, "pid": 2, "tid": 0, "ts": 7253740949588.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740949633.241, "dur": 78.820,
    "args": {
      "External id": 2526,      "device": 2, "stream": 0,      "correlation": 1274, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1702.8168635262177
    }
  },
  {
    "ph": "f", "id": 1274, "pid": 2, "tid": 0, "ts": 7253740949633.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949719.840, "dur": 5.411,
    "args": {
      "External id": 2533,      "device": 2, "stream": 0,      "correlation": 1275, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1275, "pid": 2, "tid": 0, "ts": 7253740949719.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949725.252, "dur": 5.572,
    "args": {
      "External id": 2534,      "device": 2, "stream": 0,      "correlation": 1276, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1276, "pid": 2, "tid": 0, "ts": 7253740949725.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740949730.825, "dur": 84.112,
    "args": {
      "External id": 2535,      "device": 2, "stream": 0,      "correlation": 1277, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1277, "pid": 2, "tid": 0, "ts": 7253740949730.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949814.938, "dur": 63.385,
    "args": {
      "External id": 2536,      "device": 2, "stream": 0,      "correlation": 1278, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1278, "pid": 2, "tid": 0, "ts": 7253740949814.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740949878.324, "dur": 76.855,
    "args": {
      "External id": 2537,      "device": 2, "stream": 0,      "correlation": 1279, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1279, "pid": 2, "tid": 0, "ts": 7253740949878.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740949955.180, "dur": 61.060,
    "args": {
      "External id": 2538,      "device": 2, "stream": 0,      "correlation": 1280, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1280, "pid": 2, "tid": 0, "ts": 7253740949955.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740950016.241, "dur": 74.570,
    "args": {
      "External id": 2540,      "device": 2, "stream": 0,      "correlation": 1281, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1281, "pid": 2, "tid": 0, "ts": 7253740950016.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740950090.812, "dur": 81.667,
    "args": {
      "External id": 2541,      "device": 2, "stream": 0,      "correlation": 1282, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1282, "pid": 2, "tid": 0, "ts": 7253740950090.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950179.816, "dur": 46.586,
    "args": {
      "External id": 2543,      "device": 2, "stream": 0,      "correlation": 1283, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1283, "pid": 2, "tid": 0, "ts": 7253740950179.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740950226.403, "dur": 77.858,
    "args": {
      "External id": 2546,      "device": 2, "stream": 0,      "correlation": 1284, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1723.856304345034
    }
  },
  {
    "ph": "f", "id": 1284, "pid": 2, "tid": 0, "ts": 7253740950226.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950311.880, "dur": 5.451,
    "args": {
      "External id": 2553,      "device": 2, "stream": 0,      "correlation": 1285, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1285, "pid": 2, "tid": 0, "ts": 7253740950311.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950317.332, "dur": 5.813,
    "args": {
      "External id": 2554,      "device": 2, "stream": 0,      "correlation": 1286, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1286, "pid": 2, "tid": 0, "ts": 7253740950317.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740950323.146, "dur": 81.666,
    "args": {
      "External id": 2555,      "device": 2, "stream": 0,      "correlation": 1287, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1287, "pid": 2, "tid": 0, "ts": 7253740950323.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950404.813, "dur": 64.267,
    "args": {
      "External id": 2556,      "device": 2, "stream": 0,      "correlation": 1288, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1288, "pid": 2, "tid": 0, "ts": 7253740950404.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740950469.081, "dur": 76.976,
    "args": {
      "External id": 2557,      "device": 2, "stream": 0,      "correlation": 1289, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1289, "pid": 2, "tid": 0, "ts": 7253740950469.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950546.058, "dur": 63.144,
    "args": {
      "External id": 2558,      "device": 2, "stream": 0,      "correlation": 1290, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1290, "pid": 2, "tid": 0, "ts": 7253740950546.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740950609.203, "dur": 74.049,
    "args": {
      "External id": 2560,      "device": 2, "stream": 0,      "correlation": 1291, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1291, "pid": 2, "tid": 0, "ts": 7253740950609.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740950683.253, "dur": 81.987,
    "args": {
      "External id": 2561,      "device": 2, "stream": 0,      "correlation": 1292, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1292, "pid": 2, "tid": 0, "ts": 7253740950683.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950772.097, "dur": 44.621,
    "args": {
      "External id": 2563,      "device": 2, "stream": 0,      "correlation": 1293, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1293, "pid": 2, "tid": 0, "ts": 7253740950772.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740950816.719, "dur": 77.177,
    "args": {
      "External id": 2566,      "device": 2, "stream": 0,      "correlation": 1294, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1739.067195314727
    }
  },
  {
    "ph": "f", "id": 1294, "pid": 2, "tid": 0, "ts": 7253740950816.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950901.274, "dur": 5.251,
    "args": {
      "External id": 2573,      "device": 2, "stream": 0,      "correlation": 1295, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1295, "pid": 2, "tid": 0, "ts": 7253740950901.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950906.526, "dur": 5.772,
    "args": {
      "External id": 2574,      "device": 2, "stream": 0,      "correlation": 1296, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1296, "pid": 2, "tid": 0, "ts": 7253740950906.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740950912.299, "dur": 81.506,
    "args": {
      "External id": 2575,      "device": 2, "stream": 0,      "correlation": 1297, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1297, "pid": 2, "tid": 0, "ts": 7253740950912.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740950993.806, "dur": 63.304,
    "args": {
      "External id": 2576,      "device": 2, "stream": 0,      "correlation": 1298, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1298, "pid": 2, "tid": 0, "ts": 7253740950993.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740951057.111, "dur": 77.498,
    "args": {
      "External id": 2577,      "device": 2, "stream": 0,      "correlation": 1299, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1299, "pid": 2, "tid": 0, "ts": 7253740951057.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951134.610, "dur": 61.500,
    "args": {
      "External id": 2578,      "device": 2, "stream": 0,      "correlation": 1300, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1300, "pid": 2, "tid": 0, "ts": 7253740951134.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740951196.111, "dur": 75.051,
    "args": {
      "External id": 2580,      "device": 2, "stream": 0,      "correlation": 1301, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1301, "pid": 2, "tid": 0, "ts": 7253740951196.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740951271.163, "dur": 81.667,
    "args": {
      "External id": 2581,      "device": 2, "stream": 0,      "correlation": 1302, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1302, "pid": 2, "tid": 0, "ts": 7253740951271.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951360.127, "dur": 46.226,
    "args": {
      "External id": 2583,      "device": 2, "stream": 0,      "correlation": 1303, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1303, "pid": 2, "tid": 0, "ts": 7253740951360.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740951406.354, "dur": 77.376,
    "args": {
      "External id": 2586,      "device": 2, "stream": 0,      "correlation": 1304, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1734.594621140649
    }
  },
  {
    "ph": "f", "id": 1304, "pid": 2, "tid": 0, "ts": 7253740951406.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951491.229, "dur": 5.331,
    "args": {
      "External id": 2593,      "device": 2, "stream": 0,      "correlation": 1305, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1305, "pid": 2, "tid": 0, "ts": 7253740951491.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951496.561, "dur": 5.451,
    "args": {
      "External id": 2594,      "device": 2, "stream": 0,      "correlation": 1306, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1306, "pid": 2, "tid": 0, "ts": 7253740951496.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740951502.013, "dur": 81.106,
    "args": {
      "External id": 2595,      "device": 2, "stream": 0,      "correlation": 1307, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1307, "pid": 2, "tid": 0, "ts": 7253740951502.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951583.120, "dur": 61.700,
    "args": {
      "External id": 2596,      "device": 2, "stream": 0,      "correlation": 1308, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1308, "pid": 2, "tid": 0, "ts": 7253740951583.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740951644.821, "dur": 76.295,
    "args": {
      "External id": 2597,      "device": 2, "stream": 0,      "correlation": 1309, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1309, "pid": 2, "tid": 0, "ts": 7253740951644.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951721.117, "dur": 62.943,
    "args": {
      "External id": 2598,      "device": 2, "stream": 0,      "correlation": 1310, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1310, "pid": 2, "tid": 0, "ts": 7253740951721.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740951784.061, "dur": 75.693,
    "args": {
      "External id": 2600,      "device": 2, "stream": 0,      "correlation": 1311, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1311, "pid": 2, "tid": 0, "ts": 7253740951784.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740951859.755, "dur": 81.105,
    "args": {
      "External id": 2601,      "device": 2, "stream": 0,      "correlation": 1312, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1312, "pid": 2, "tid": 0, "ts": 7253740951859.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740951948.399, "dur": 46.385,
    "args": {
      "External id": 2603,      "device": 2, "stream": 0,      "correlation": 1313, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1313, "pid": 2, "tid": 0, "ts": 7253740951948.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740951994.785, "dur": 79.943,
    "args": {
      "External id": 2606,      "device": 2, "stream": 0,      "correlation": 1314, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1678.8968277794456
    }
  },
  {
    "ph": "f", "id": 1314, "pid": 2, "tid": 0, "ts": 7253740951994.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952082.827, "dur": 5.492,
    "args": {
      "External id": 2613,      "device": 2, "stream": 0,      "correlation": 1315, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1315, "pid": 2, "tid": 0, "ts": 7253740952082.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952088.320, "dur": 6.133,
    "args": {
      "External id": 2614,      "device": 2, "stream": 0,      "correlation": 1316, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1316, "pid": 2, "tid": 0, "ts": 7253740952088.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740952094.454, "dur": 82.148,
    "args": {
      "External id": 2615,      "device": 2, "stream": 0,      "correlation": 1317, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1317, "pid": 2, "tid": 0, "ts": 7253740952094.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952176.603, "dur": 64.667,
    "args": {
      "External id": 2616,      "device": 2, "stream": 0,      "correlation": 1318, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1318, "pid": 2, "tid": 0, "ts": 7253740952176.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740952241.271, "dur": 75.814,
    "args": {
      "External id": 2617,      "device": 2, "stream": 0,      "correlation": 1319, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1319, "pid": 2, "tid": 0, "ts": 7253740952241.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952317.086, "dur": 64.868,
    "args": {
      "External id": 2618,      "device": 2, "stream": 0,      "correlation": 1320, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1320, "pid": 2, "tid": 0, "ts": 7253740952317.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740952381.955, "dur": 75.211,
    "args": {
      "External id": 2620,      "device": 2, "stream": 0,      "correlation": 1321, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1321, "pid": 2, "tid": 0, "ts": 7253740952381.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740952457.167, "dur": 81.466,
    "args": {
      "External id": 2621,      "device": 2, "stream": 0,      "correlation": 1322, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1322, "pid": 2, "tid": 0, "ts": 7253740952457.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952545.811, "dur": 45.183,
    "args": {
      "External id": 2623,      "device": 2, "stream": 0,      "correlation": 1323, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1323, "pid": 2, "tid": 0, "ts": 7253740952545.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740952590.995, "dur": 78.539,
    "args": {
      "External id": 2626,      "device": 2, "stream": 0,      "correlation": 1324, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1708.9091927680163
    }
  },
  {
    "ph": "f", "id": 1324, "pid": 2, "tid": 0, "ts": 7253740952590.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952676.872, "dur": 5.572,
    "args": {
      "External id": 2633,      "device": 2, "stream": 0,      "correlation": 1325, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1325, "pid": 2, "tid": 0, "ts": 7253740952676.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952682.445, "dur": 6.413,
    "args": {
      "External id": 2634,      "device": 2, "stream": 0,      "correlation": 1326, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1326, "pid": 2, "tid": 0, "ts": 7253740952682.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740952688.859, "dur": 82.108,
    "args": {
      "External id": 2635,      "device": 2, "stream": 0,      "correlation": 1327, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1327, "pid": 2, "tid": 0, "ts": 7253740952688.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952770.968, "dur": 61.901,
    "args": {
      "External id": 2636,      "device": 2, "stream": 0,      "correlation": 1328, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1328, "pid": 2, "tid": 0, "ts": 7253740952770.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740952832.870, "dur": 77.337,
    "args": {
      "External id": 2637,      "device": 2, "stream": 0,      "correlation": 1329, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1329, "pid": 2, "tid": 0, "ts": 7253740952832.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740952910.208, "dur": 63.785,
    "args": {
      "External id": 2638,      "device": 2, "stream": 0,      "correlation": 1330, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1330, "pid": 2, "tid": 0, "ts": 7253740952910.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740952973.994, "dur": 73.007,
    "args": {
      "External id": 2640,      "device": 2, "stream": 0,      "correlation": 1331, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1331, "pid": 2, "tid": 0, "ts": 7253740952973.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740953047.002, "dur": 80.584,
    "args": {
      "External id": 2641,      "device": 2, "stream": 0,      "correlation": 1332, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1332, "pid": 2, "tid": 0, "ts": 7253740953047.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953135.606, "dur": 44.982,
    "args": {
      "External id": 2643,      "device": 2, "stream": 0,      "correlation": 1333, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1333, "pid": 2, "tid": 0, "ts": 7253740953135.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740953180.589, "dur": 78.379,
    "args": {
      "External id": 2646,      "device": 2, "stream": 0,      "correlation": 1334, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1712.3976524623629
    }
  },
  {
    "ph": "f", "id": 1334, "pid": 2, "tid": 0, "ts": 7253740953180.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953265.985, "dur": 5.652,
    "args": {
      "External id": 2653,      "device": 2, "stream": 0,      "correlation": 1335, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1335, "pid": 2, "tid": 0, "ts": 7253740953265.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953271.638, "dur": 5.933,
    "args": {
      "External id": 2654,      "device": 2, "stream": 0,      "correlation": 1336, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1336, "pid": 2, "tid": 0, "ts": 7253740953271.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740953277.572, "dur": 80.504,
    "args": {
      "External id": 2655,      "device": 2, "stream": 0,      "correlation": 1337, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1337, "pid": 2, "tid": 0, "ts": 7253740953277.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953358.077, "dur": 63.825,
    "args": {
      "External id": 2656,      "device": 2, "stream": 0,      "correlation": 1338, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1338, "pid": 2, "tid": 0, "ts": 7253740953358.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740953421.903, "dur": 77.056,
    "args": {
      "External id": 2657,      "device": 2, "stream": 0,      "correlation": 1339, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1339, "pid": 2, "tid": 0, "ts": 7253740953421.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953498.960, "dur": 61.621,
    "args": {
      "External id": 2658,      "device": 2, "stream": 0,      "correlation": 1340, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1340, "pid": 2, "tid": 0, "ts": 7253740953498.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740953560.582, "dur": 74.329,
    "args": {
      "External id": 2660,      "device": 2, "stream": 0,      "correlation": 1341, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1341, "pid": 2, "tid": 0, "ts": 7253740953560.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740953634.912, "dur": 79.903,
    "args": {
      "External id": 2661,      "device": 2, "stream": 0,      "correlation": 1342, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1342, "pid": 2, "tid": 0, "ts": 7253740953634.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953722.834, "dur": 45.504,
    "args": {
      "External id": 2663,      "device": 2, "stream": 0,      "correlation": 1343, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1343, "pid": 2, "tid": 0, "ts": 7253740953722.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740953768.339, "dur": 76.374,
    "args": {
      "External id": 2666,      "device": 2, "stream": 0,      "correlation": 1344, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1757.3515941080198
    }
  },
  {
    "ph": "f", "id": 1344, "pid": 2, "tid": 0, "ts": 7253740953768.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953852.653, "dur": 5.491,
    "args": {
      "External id": 2673,      "device": 2, "stream": 0,      "correlation": 1345, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1345, "pid": 2, "tid": 0, "ts": 7253740953852.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953858.145, "dur": 5.933,
    "args": {
      "External id": 2674,      "device": 2, "stream": 0,      "correlation": 1346, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1346, "pid": 2, "tid": 0, "ts": 7253740953858.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740953864.079, "dur": 81.265,
    "args": {
      "External id": 2675,      "device": 2, "stream": 0,      "correlation": 1347, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1347, "pid": 2, "tid": 0, "ts": 7253740953864.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740953945.345, "dur": 71.363,
    "args": {
      "External id": 2676,      "device": 2, "stream": 0,      "correlation": 1348, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1348, "pid": 2, "tid": 0, "ts": 7253740953945.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740954016.709, "dur": 77.217,
    "args": {
      "External id": 2677,      "device": 2, "stream": 0,      "correlation": 1349, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1349, "pid": 2, "tid": 0, "ts": 7253740954016.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954093.927, "dur": 63.745,
    "args": {
      "External id": 2678,      "device": 2, "stream": 0,      "correlation": 1350, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1350, "pid": 2, "tid": 0, "ts": 7253740954093.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740954157.673, "dur": 76.134,
    "args": {
      "External id": 2680,      "device": 2, "stream": 0,      "correlation": 1351, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1351, "pid": 2, "tid": 0, "ts": 7253740954157.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740954233.808, "dur": 82.228,
    "args": {
      "External id": 2681,      "device": 2, "stream": 0,      "correlation": 1352, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1352, "pid": 2, "tid": 0, "ts": 7253740954233.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954323.213, "dur": 44.943,
    "args": {
      "External id": 2683,      "device": 2, "stream": 0,      "correlation": 1353, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1353, "pid": 2, "tid": 0, "ts": 7253740954323.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740954368.157, "dur": 77.096,
    "args": {
      "External id": 2686,      "device": 2, "stream": 0,      "correlation": 1354, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1740.8943019832159
    }
  },
  {
    "ph": "f", "id": 1354, "pid": 2, "tid": 0, "ts": 7253740954368.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954453.272, "dur": 5.331,
    "args": {
      "External id": 2693,      "device": 2, "stream": 0,      "correlation": 1355, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1355, "pid": 2, "tid": 0, "ts": 7253740954453.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954458.604, "dur": 5.853,
    "args": {
      "External id": 2694,      "device": 2, "stream": 0,      "correlation": 1356, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1356, "pid": 2, "tid": 0, "ts": 7253740954458.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740954464.458, "dur": 82.468,
    "args": {
      "External id": 2695,      "device": 2, "stream": 0,      "correlation": 1357, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1357, "pid": 2, "tid": 0, "ts": 7253740954464.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954546.927, "dur": 60.699,
    "args": {
      "External id": 2696,      "device": 2, "stream": 0,      "correlation": 1358, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1358, "pid": 2, "tid": 0, "ts": 7253740954546.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740954607.627, "dur": 78.419,
    "args": {
      "External id": 2697,      "device": 2, "stream": 0,      "correlation": 1359, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1359, "pid": 2, "tid": 0, "ts": 7253740954607.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954686.047, "dur": 64.547,
    "args": {
      "External id": 2698,      "device": 2, "stream": 0,      "correlation": 1360, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1360, "pid": 2, "tid": 0, "ts": 7253740954686.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740954750.595, "dur": 74.691,
    "args": {
      "External id": 2700,      "device": 2, "stream": 0,      "correlation": 1361, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1361, "pid": 2, "tid": 0, "ts": 7253740954750.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740954825.287, "dur": 82.227,
    "args": {
      "External id": 2701,      "device": 2, "stream": 0,      "correlation": 1362, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1362, "pid": 2, "tid": 0, "ts": 7253740954825.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740954915.093, "dur": 43.980,
    "args": {
      "External id": 2703,      "device": 2, "stream": 0,      "correlation": 1363, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1363, "pid": 2, "tid": 0, "ts": 7253740954915.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740954959.074, "dur": 77.737,
    "args": {
      "External id": 2706,      "device": 2, "stream": 0,      "correlation": 1364, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1726.539504489439
    }
  },
  {
    "ph": "f", "id": 1364, "pid": 2, "tid": 0, "ts": 7253740954959.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955044.109, "dur": 5.331,
    "args": {
      "External id": 2713,      "device": 2, "stream": 0,      "correlation": 1365, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1365, "pid": 2, "tid": 0, "ts": 7253740955044.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955049.441, "dur": 5.933,
    "args": {
      "External id": 2714,      "device": 2, "stream": 0,      "correlation": 1366, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1366, "pid": 2, "tid": 0, "ts": 7253740955049.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740955055.375, "dur": 82.669,
    "args": {
      "External id": 2715,      "device": 2, "stream": 0,      "correlation": 1367, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1367, "pid": 2, "tid": 0, "ts": 7253740955055.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955138.045, "dur": 64.267,
    "args": {
      "External id": 2716,      "device": 2, "stream": 0,      "correlation": 1368, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1368, "pid": 2, "tid": 0, "ts": 7253740955138.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740955202.313, "dur": 77.577,
    "args": {
      "External id": 2717,      "device": 2, "stream": 0,      "correlation": 1369, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1369, "pid": 2, "tid": 0, "ts": 7253740955202.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955279.891, "dur": 60.297,
    "args": {
      "External id": 2718,      "device": 2, "stream": 0,      "correlation": 1370, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1370, "pid": 2, "tid": 0, "ts": 7253740955279.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740955340.189, "dur": 76.134,
    "args": {
      "External id": 2720,      "device": 2, "stream": 0,      "correlation": 1371, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1371, "pid": 2, "tid": 0, "ts": 7253740955340.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740955416.324, "dur": 82.228,
    "args": {
      "External id": 2721,      "device": 2, "stream": 0,      "correlation": 1372, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1372, "pid": 2, "tid": 0, "ts": 7253740955416.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955506.251, "dur": 44.822,
    "args": {
      "External id": 2723,      "device": 2, "stream": 0,      "correlation": 1373, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1373, "pid": 2, "tid": 0, "ts": 7253740955506.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740955551.074, "dur": 77.457,
    "args": {
      "External id": 2726,      "device": 2, "stream": 0,      "correlation": 1374, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1732.7807069637738
    }
  },
  {
    "ph": "f", "id": 1374, "pid": 2, "tid": 0, "ts": 7253740955551.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955636.069, "dur": 5.411,
    "args": {
      "External id": 2733,      "device": 2, "stream": 0,      "correlation": 1375, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1375, "pid": 2, "tid": 0, "ts": 7253740955636.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955641.481, "dur": 6.133,
    "args": {
      "External id": 2734,      "device": 2, "stream": 0,      "correlation": 1376, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1376, "pid": 2, "tid": 0, "ts": 7253740955641.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740955647.615, "dur": 82.709,
    "args": {
      "External id": 2735,      "device": 2, "stream": 0,      "correlation": 1377, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1377, "pid": 2, "tid": 0, "ts": 7253740955647.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955730.325, "dur": 62.222,
    "args": {
      "External id": 2736,      "device": 2, "stream": 0,      "correlation": 1378, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1378, "pid": 2, "tid": 0, "ts": 7253740955730.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740955792.548, "dur": 78.500,
    "args": {
      "External id": 2737,      "device": 2, "stream": 0,      "correlation": 1379, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1379, "pid": 2, "tid": 0, "ts": 7253740955792.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740955871.049, "dur": 66.070,
    "args": {
      "External id": 2738,      "device": 2, "stream": 0,      "correlation": 1380, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1380, "pid": 2, "tid": 0, "ts": 7253740955871.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740955937.120, "dur": 75.653,
    "args": {
      "External id": 2740,      "device": 2, "stream": 0,      "correlation": 1381, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1381, "pid": 2, "tid": 0, "ts": 7253740955937.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740956012.774, "dur": 82.829,
    "args": {
      "External id": 2741,      "device": 2, "stream": 0,      "correlation": 1382, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1382, "pid": 2, "tid": 0, "ts": 7253740956012.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956103.142, "dur": 45.583,
    "args": {
      "External id": 2743,      "device": 2, "stream": 0,      "correlation": 1383, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1383, "pid": 2, "tid": 0, "ts": 7253740956103.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740956148.726, "dur": 78.620,
    "args": {
      "External id": 2746,      "device": 2, "stream": 0,      "correlation": 1384, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1707.1485735363326
    }
  },
  {
    "ph": "f", "id": 1384, "pid": 2, "tid": 0, "ts": 7253740956148.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956234.764, "dur": 5.732,
    "args": {
      "External id": 2753,      "device": 2, "stream": 0,      "correlation": 1385, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1385, "pid": 2, "tid": 0, "ts": 7253740956234.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956240.497, "dur": 6.053,
    "args": {
      "External id": 2754,      "device": 2, "stream": 0,      "correlation": 1386, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1386, "pid": 2, "tid": 0, "ts": 7253740956240.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740956246.551, "dur": 82.629,
    "args": {
      "External id": 2755,      "device": 2, "stream": 0,      "correlation": 1387, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1387, "pid": 2, "tid": 0, "ts": 7253740956246.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956329.181, "dur": 62.101,
    "args": {
      "External id": 2756,      "device": 2, "stream": 0,      "correlation": 1388, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1388, "pid": 2, "tid": 0, "ts": 7253740956329.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740956391.283, "dur": 75.934,
    "args": {
      "External id": 2757,      "device": 2, "stream": 0,      "correlation": 1389, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1389, "pid": 2, "tid": 0, "ts": 7253740956391.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956467.218, "dur": 67.955,
    "args": {
      "External id": 2758,      "device": 2, "stream": 0,      "correlation": 1390, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1390, "pid": 2, "tid": 0, "ts": 7253740956467.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740956535.174, "dur": 75.733,
    "args": {
      "External id": 2760,      "device": 2, "stream": 0,      "correlation": 1391, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1391, "pid": 2, "tid": 0, "ts": 7253740956535.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740956610.908, "dur": 82.148,
    "args": {
      "External id": 2761,      "device": 2, "stream": 0,      "correlation": 1392, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1392, "pid": 2, "tid": 0, "ts": 7253740956610.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956701.356, "dur": 45.223,
    "args": {
      "External id": 2763,      "device": 2, "stream": 0,      "correlation": 1393, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1393, "pid": 2, "tid": 0, "ts": 7253740956701.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740956746.580, "dur": 79.581,
    "args": {
      "External id": 2766,      "device": 2, "stream": 0,      "correlation": 1394, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1686.5337387851523
    }
  },
  {
    "ph": "f", "id": 1394, "pid": 2, "tid": 0, "ts": 7253740956746.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956833.579, "dur": 5.452,
    "args": {
      "External id": 2773,      "device": 2, "stream": 0,      "correlation": 1395, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1395, "pid": 2, "tid": 0, "ts": 7253740956833.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956839.032, "dur": 5.852,
    "args": {
      "External id": 2774,      "device": 2, "stream": 0,      "correlation": 1396, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1396, "pid": 2, "tid": 0, "ts": 7253740956839.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740956844.885, "dur": 82.268,
    "args": {
      "External id": 2775,      "device": 2, "stream": 0,      "correlation": 1397, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1397, "pid": 2, "tid": 0, "ts": 7253740956844.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740956927.154, "dur": 62.262,
    "args": {
      "External id": 2776,      "device": 2, "stream": 0,      "correlation": 1398, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1398, "pid": 2, "tid": 0, "ts": 7253740956927.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740956989.417, "dur": 75.934,
    "args": {
      "External id": 2777,      "device": 2, "stream": 0,      "correlation": 1399, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1399, "pid": 2, "tid": 0, "ts": 7253740956989.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957065.352, "dur": 63.625,
    "args": {
      "External id": 2778,      "device": 2, "stream": 0,      "correlation": 1400, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1400, "pid": 2, "tid": 0, "ts": 7253740957065.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740957128.978, "dur": 75.252,
    "args": {
      "External id": 2780,      "device": 2, "stream": 0,      "correlation": 1401, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1401, "pid": 2, "tid": 0, "ts": 7253740957128.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740957204.231, "dur": 81.626,
    "args": {
      "External id": 2781,      "device": 2, "stream": 0,      "correlation": 1402, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1402, "pid": 2, "tid": 0, "ts": 7253740957204.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957293.476, "dur": 44.902,
    "args": {
      "External id": 2783,      "device": 2, "stream": 0,      "correlation": 1403, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1403, "pid": 2, "tid": 0, "ts": 7253740957293.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740957338.379, "dur": 76.414,
    "args": {
      "External id": 2786,      "device": 2, "stream": 0,      "correlation": 1404, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1756.4316953477721
    }
  },
  {
    "ph": "f", "id": 1404, "pid": 2, "tid": 0, "ts": 7253740957338.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957422.612, "dur": 5.291,
    "args": {
      "External id": 2793,      "device": 2, "stream": 0,      "correlation": 1405, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1405, "pid": 2, "tid": 0, "ts": 7253740957422.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957427.904, "dur": 6.053,
    "args": {
      "External id": 2794,      "device": 2, "stream": 0,      "correlation": 1406, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1406, "pid": 2, "tid": 0, "ts": 7253740957427.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740957433.958, "dur": 81.386,
    "args": {
      "External id": 2795,      "device": 2, "stream": 0,      "correlation": 1407, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1407, "pid": 2, "tid": 0, "ts": 7253740957433.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957515.345, "dur": 63.184,
    "args": {
      "External id": 2796,      "device": 2, "stream": 0,      "correlation": 1408, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1408, "pid": 2, "tid": 0, "ts": 7253740957515.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740957578.530, "dur": 76.094,
    "args": {
      "External id": 2797,      "device": 2, "stream": 0,      "correlation": 1409, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1409, "pid": 2, "tid": 0, "ts": 7253740957578.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957654.625, "dur": 65.871,
    "args": {
      "External id": 2798,      "device": 2, "stream": 0,      "correlation": 1410, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1410, "pid": 2, "tid": 0, "ts": 7253740957654.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740957720.497, "dur": 74.931,
    "args": {
      "External id": 2800,      "device": 2, "stream": 0,      "correlation": 1411, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1411, "pid": 2, "tid": 0, "ts": 7253740957720.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740957795.429, "dur": 80.864,
    "args": {
      "External id": 2801,      "device": 2, "stream": 0,      "correlation": 1412, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1412, "pid": 2, "tid": 0, "ts": 7253740957795.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740957883.230, "dur": 44.943,
    "args": {
      "External id": 2803,      "device": 2, "stream": 0,      "correlation": 1413, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1413, "pid": 2, "tid": 0, "ts": 7253740957883.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740957928.174, "dur": 77.577,
    "args": {
      "External id": 2806,      "device": 2, "stream": 0,      "correlation": 1414, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1730.1003892856222
    }
  },
  {
    "ph": "f", "id": 1414, "pid": 2, "tid": 0, "ts": 7253740957928.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958012.728, "dur": 5.491,
    "args": {
      "External id": 2813,      "device": 2, "stream": 0,      "correlation": 1415, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1415, "pid": 2, "tid": 0, "ts": 7253740958012.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958018.220, "dur": 6.334,
    "args": {
      "External id": 2814,      "device": 2, "stream": 0,      "correlation": 1416, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1416, "pid": 2, "tid": 0, "ts": 7253740958018.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740958024.555, "dur": 80.744,
    "args": {
      "External id": 2815,      "device": 2, "stream": 0,      "correlation": 1417, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1417, "pid": 2, "tid": 0, "ts": 7253740958024.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958105.300, "dur": 61.701,
    "args": {
      "External id": 2816,      "device": 2, "stream": 0,      "correlation": 1418, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1418, "pid": 2, "tid": 0, "ts": 7253740958105.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740958167.002, "dur": 78.700,
    "args": {
      "External id": 2817,      "device": 2, "stream": 0,      "correlation": 1419, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1419, "pid": 2, "tid": 0, "ts": 7253740958167.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958245.703, "dur": 63.184,
    "args": {
      "External id": 2818,      "device": 2, "stream": 0,      "correlation": 1420, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1420, "pid": 2, "tid": 0, "ts": 7253740958245.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740958308.888, "dur": 74.931,
    "args": {
      "External id": 2820,      "device": 2, "stream": 0,      "correlation": 1421, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1421, "pid": 2, "tid": 0, "ts": 7253740958308.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740958383.820, "dur": 80.945,
    "args": {
      "External id": 2821,      "device": 2, "stream": 0,      "correlation": 1422, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1422, "pid": 2, "tid": 0, "ts": 7253740958383.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958472.744, "dur": 45.143,
    "args": {
      "External id": 2823,      "device": 2, "stream": 0,      "correlation": 1423, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1423, "pid": 2, "tid": 0, "ts": 7253740958472.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740958517.888, "dur": 80.704,
    "args": {
      "External id": 2826,      "device": 2, "stream": 0,      "correlation": 1424, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1663.0658323523946
    }
  },
  {
    "ph": "f", "id": 1424, "pid": 2, "tid": 0, "ts": 7253740958517.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958606.532, "dur": 5.411,
    "args": {
      "External id": 2833,      "device": 2, "stream": 0,      "correlation": 1425, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1425, "pid": 2, "tid": 0, "ts": 7253740958606.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958611.944, "dur": 5.612,
    "args": {
      "External id": 2834,      "device": 2, "stream": 0,      "correlation": 1426, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1426, "pid": 2, "tid": 0, "ts": 7253740958611.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740958617.557, "dur": 80.744,
    "args": {
      "External id": 2835,      "device": 2, "stream": 0,      "correlation": 1427, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1427, "pid": 2, "tid": 0, "ts": 7253740958617.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740917537.772, "dur": 43277.888,
    "args": {
            "cid": 48, "correlation": 1463
    }
  },
  {
    "ph": "f", "id": 1463, "pid": 49625, "tid": 49625, "ts": 7253740917537.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960862.482, "dur": 2.151,
    "args": {
            "cid": 57, "correlation": 1464
    }
  },
  {
    "ph": "f", "id": 1464, "pid": 49625, "tid": 49625, "ts": 7253740960862.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960868.157, "dur": 0.947,
    "args": {
            "cid": 57, "correlation": 1465
    }
  },
  {
    "ph": "f", "id": 1465, "pid": 49625, "tid": 49625, "ts": 7253740960868.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960871.944, "dur": 0.810,
    "args": {
            "cid": 57, "correlation": 1466
    }
  },
  {
    "ph": "f", "id": 1466, "pid": 49625, "tid": 49625, "ts": 7253740960871.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960876.096, "dur": 0.596,
    "args": {
            "cid": 57, "correlation": 1467
    }
  },
  {
    "ph": "f", "id": 1467, "pid": 49625, "tid": 49625, "ts": 7253740960876.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960878.482, "dur": 0.520,
    "args": {
            "cid": 57, "correlation": 1468
    }
  },
  {
    "ph": "f", "id": 1468, "pid": 49625, "tid": 49625, "ts": 7253740960878.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960880.969, "dur": 0.518,
    "args": {
            "cid": 57, "correlation": 1469
    }
  },
  {
    "ph": "f", "id": 1469, "pid": 49625, "tid": 49625, "ts": 7253740960880.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960882.837, "dur": 0.693,
    "args": {
            "cid": 57, "correlation": 1470
    }
  },
  {
    "ph": "f", "id": 1470, "pid": 49625, "tid": 49625, "ts": 7253740960882.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960884.900, "dur": 0.476,
    "args": {
            "cid": 57, "correlation": 1471
    }
  },
  {
    "ph": "f", "id": 1471, "pid": 49625, "tid": 49625, "ts": 7253740960884.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960886.678, "dur": 0.613,
    "args": {
            "cid": 57, "correlation": 1472
    }
  },
  {
    "ph": "f", "id": 1472, "pid": 49625, "tid": 49625, "ts": 7253740960886.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960888.745, "dur": 0.947,
    "args": {
            "cid": 57, "correlation": 1473
    }
  },
  {
    "ph": "f", "id": 1473, "pid": 49625, "tid": 49625, "ts": 7253740960888.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960891.234, "dur": 0.598,
    "args": {
            "cid": 57, "correlation": 1474
    }
  },
  {
    "ph": "f", "id": 1474, "pid": 49625, "tid": 49625, "ts": 7253740960891.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960893.318, "dur": 0.639,
    "args": {
            "cid": 57, "correlation": 1475
    }
  },
  {
    "ph": "f", "id": 1475, "pid": 49625, "tid": 49625, "ts": 7253740960893.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960895.177, "dur": 0.595,
    "args": {
            "cid": 57, "correlation": 1476
    }
  },
  {
    "ph": "f", "id": 1476, "pid": 49625, "tid": 49625, "ts": 7253740960895.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960897.010, "dur": 0.549,
    "args": {
            "cid": 57, "correlation": 1477
    }
  },
  {
    "ph": "f", "id": 1477, "pid": 49625, "tid": 49625, "ts": 7253740960897.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960898.941, "dur": 0.649,
    "args": {
            "cid": 57, "correlation": 1478
    }
  },
  {
    "ph": "f", "id": 1478, "pid": 49625, "tid": 49625, "ts": 7253740960898.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960900.936, "dur": 0.592,
    "args": {
            "cid": 57, "correlation": 1479
    }
  },
  {
    "ph": "f", "id": 1479, "pid": 49625, "tid": 49625, "ts": 7253740960900.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960902.835, "dur": 0.522,
    "args": {
            "cid": 57, "correlation": 1480
    }
  },
  {
    "ph": "f", "id": 1480, "pid": 49625, "tid": 49625, "ts": 7253740960902.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960904.903, "dur": 0.525,
    "args": {
            "cid": 57, "correlation": 1481
    }
  },
  {
    "ph": "f", "id": 1481, "pid": 49625, "tid": 49625, "ts": 7253740960904.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960906.706, "dur": 0.744,
    "args": {
            "cid": 57, "correlation": 1482
    }
  },
  {
    "ph": "f", "id": 1482, "pid": 49625, "tid": 49625, "ts": 7253740960906.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960908.766, "dur": 0.688,
    "args": {
            "cid": 57, "correlation": 1483
    }
  },
  {
    "ph": "f", "id": 1483, "pid": 49625, "tid": 49625, "ts": 7253740960908.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960910.828, "dur": 0.634,
    "args": {
            "cid": 57, "correlation": 1484
    }
  },
  {
    "ph": "f", "id": 1484, "pid": 49625, "tid": 49625, "ts": 7253740960910.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960912.757, "dur": 0.542,
    "args": {
            "cid": 57, "correlation": 1485
    }
  },
  {
    "ph": "f", "id": 1485, "pid": 49625, "tid": 49625, "ts": 7253740960912.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960914.472, "dur": 0.448,
    "args": {
            "cid": 57, "correlation": 1486
    }
  },
  {
    "ph": "f", "id": 1486, "pid": 49625, "tid": 49625, "ts": 7253740960914.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960916.259, "dur": 0.703,
    "args": {
            "cid": 57, "correlation": 1487
    }
  },
  {
    "ph": "f", "id": 1487, "pid": 49625, "tid": 49625, "ts": 7253740960916.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960918.316, "dur": 0.454,
    "args": {
            "cid": 57, "correlation": 1488
    }
  },
  {
    "ph": "f", "id": 1488, "pid": 49625, "tid": 49625, "ts": 7253740960918.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960920.183, "dur": 0.582,
    "args": {
            "cid": 57, "correlation": 1489
    }
  },
  {
    "ph": "f", "id": 1489, "pid": 49625, "tid": 49625, "ts": 7253740960920.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960922.093, "dur": 0.713,
    "args": {
            "cid": 57, "correlation": 1490
    }
  },
  {
    "ph": "f", "id": 1490, "pid": 49625, "tid": 49625, "ts": 7253740960922.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960924.054, "dur": 0.472,
    "args": {
            "cid": 57, "correlation": 1491
    }
  },
  {
    "ph": "f", "id": 1491, "pid": 49625, "tid": 49625, "ts": 7253740960924.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960925.900, "dur": 0.623,
    "args": {
            "cid": 57, "correlation": 1492
    }
  },
  {
    "ph": "f", "id": 1492, "pid": 49625, "tid": 49625, "ts": 7253740960925.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960927.867, "dur": 0.761,
    "args": {
            "cid": 57, "correlation": 1493
    }
  },
  {
    "ph": "f", "id": 1493, "pid": 49625, "tid": 49625, "ts": 7253740960927.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960930.100, "dur": 0.495,
    "args": {
            "cid": 57, "correlation": 1494
    }
  },
  {
    "ph": "f", "id": 1494, "pid": 49625, "tid": 49625, "ts": 7253740960930.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960931.923, "dur": 1.010,
    "args": {
            "cid": 57, "correlation": 1495
    }
  },
  {
    "ph": "f", "id": 1495, "pid": 49625, "tid": 49625, "ts": 7253740960931.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960934.157, "dur": 0.698,
    "args": {
            "cid": 57, "correlation": 1496
    }
  },
  {
    "ph": "f", "id": 1496, "pid": 49625, "tid": 49625, "ts": 7253740960934.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960936.628, "dur": 0.571,
    "args": {
            "cid": 57, "correlation": 1497
    }
  },
  {
    "ph": "f", "id": 1497, "pid": 49625, "tid": 49625, "ts": 7253740960936.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960938.700, "dur": 0.667,
    "args": {
            "cid": 57, "correlation": 1498
    }
  },
  {
    "ph": "f", "id": 1498, "pid": 49625, "tid": 49625, "ts": 7253740960938.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960940.571, "dur": 0.521,
    "args": {
            "cid": 57, "correlation": 1499
    }
  },
  {
    "ph": "f", "id": 1499, "pid": 49625, "tid": 49625, "ts": 7253740960940.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960942.401, "dur": 0.663,
    "args": {
            "cid": 57, "correlation": 1500
    }
  },
  {
    "ph": "f", "id": 1500, "pid": 49625, "tid": 49625, "ts": 7253740960942.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960944.272, "dur": 1.030,
    "args": {
            "cid": 57, "correlation": 1501
    }
  },
  {
    "ph": "f", "id": 1501, "pid": 49625, "tid": 49625, "ts": 7253740960944.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960946.605, "dur": 0.668,
    "args": {
            "cid": 57, "correlation": 1502
    }
  },
  {
    "ph": "f", "id": 1502, "pid": 49625, "tid": 49625, "ts": 7253740960946.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960948.514, "dur": 0.512,
    "args": {
            "cid": 57, "correlation": 1503
    }
  },
  {
    "ph": "f", "id": 1503, "pid": 49625, "tid": 49625, "ts": 7253740960948.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960950.317, "dur": 0.854,
    "args": {
            "cid": 57, "correlation": 1504
    }
  },
  {
    "ph": "f", "id": 1504, "pid": 49625, "tid": 49625, "ts": 7253740960950.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960952.707, "dur": 0.830,
    "args": {
            "cid": 57, "correlation": 1505
    }
  },
  {
    "ph": "f", "id": 1505, "pid": 49625, "tid": 49625, "ts": 7253740960952.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960954.834, "dur": 0.570,
    "args": {
            "cid": 57, "correlation": 1506
    }
  },
  {
    "ph": "f", "id": 1506, "pid": 49625, "tid": 49625, "ts": 7253740960954.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960956.810, "dur": 0.642,
    "args": {
            "cid": 57, "correlation": 1507
    }
  },
  {
    "ph": "f", "id": 1507, "pid": 49625, "tid": 49625, "ts": 7253740960956.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960958.829, "dur": 0.555,
    "args": {
            "cid": 57, "correlation": 1508
    }
  },
  {
    "ph": "f", "id": 1508, "pid": 49625, "tid": 49625, "ts": 7253740960958.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960960.801, "dur": 0.845,
    "args": {
            "cid": 57, "correlation": 1509
    }
  },
  {
    "ph": "f", "id": 1509, "pid": 49625, "tid": 49625, "ts": 7253740960960.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960963.031, "dur": 0.992,
    "args": {
            "cid": 57, "correlation": 1510
    }
  },
  {
    "ph": "f", "id": 1510, "pid": 49625, "tid": 49625, "ts": 7253740960963.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960965.207, "dur": 0.807,
    "args": {
            "cid": 57, "correlation": 1511
    }
  },
  {
    "ph": "f", "id": 1511, "pid": 49625, "tid": 49625, "ts": 7253740960965.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960967.378, "dur": 0.555,
    "args": {
            "cid": 57, "correlation": 1512
    }
  },
  {
    "ph": "f", "id": 1512, "pid": 49625, "tid": 49625, "ts": 7253740960967.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960969.166, "dur": 0.489,
    "args": {
            "cid": 57, "correlation": 1513
    }
  },
  {
    "ph": "f", "id": 1513, "pid": 49625, "tid": 49625, "ts": 7253740960969.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960970.949, "dur": 0.518,
    "args": {
            "cid": 57, "correlation": 1514
    }
  },
  {
    "ph": "f", "id": 1514, "pid": 49625, "tid": 49625, "ts": 7253740960970.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960972.774, "dur": 0.551,
    "args": {
            "cid": 57, "correlation": 1515
    }
  },
  {
    "ph": "f", "id": 1515, "pid": 49625, "tid": 49625, "ts": 7253740960972.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960974.506, "dur": 0.651,
    "args": {
            "cid": 57, "correlation": 1516
    }
  },
  {
    "ph": "f", "id": 1516, "pid": 49625, "tid": 49625, "ts": 7253740960974.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960977.047, "dur": 0.589,
    "args": {
            "cid": 57, "correlation": 1517
    }
  },
  {
    "ph": "f", "id": 1517, "pid": 49625, "tid": 49625, "ts": 7253740960977.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960978.850, "dur": 0.610,
    "args": {
            "cid": 57, "correlation": 1518
    }
  },
  {
    "ph": "f", "id": 1518, "pid": 49625, "tid": 49625, "ts": 7253740960978.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960980.836, "dur": 0.617,
    "args": {
            "cid": 57, "correlation": 1519
    }
  },
  {
    "ph": "f", "id": 1519, "pid": 49625, "tid": 49625, "ts": 7253740960980.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960982.734, "dur": 0.538,
    "args": {
            "cid": 57, "correlation": 1520
    }
  },
  {
    "ph": "f", "id": 1520, "pid": 49625, "tid": 49625, "ts": 7253740960982.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960984.526, "dur": 0.705,
    "args": {
            "cid": 57, "correlation": 1521
    }
  },
  {
    "ph": "f", "id": 1521, "pid": 49625, "tid": 49625, "ts": 7253740960984.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960986.435, "dur": 0.601,
    "args": {
            "cid": 57, "correlation": 1522
    }
  },
  {
    "ph": "f", "id": 1522, "pid": 49625, "tid": 49625, "ts": 7253740960986.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960988.252, "dur": 0.797,
    "args": {
            "cid": 57, "correlation": 1523
    }
  },
  {
    "ph": "f", "id": 1523, "pid": 49625, "tid": 49625, "ts": 7253740960988.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960990.363, "dur": 0.640,
    "args": {
            "cid": 57, "correlation": 1524
    }
  },
  {
    "ph": "f", "id": 1524, "pid": 49625, "tid": 49625, "ts": 7253740960990.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960992.173, "dur": 0.787,
    "args": {
            "cid": 57, "correlation": 1525
    }
  },
  {
    "ph": "f", "id": 1525, "pid": 49625, "tid": 49625, "ts": 7253740960992.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960994.208, "dur": 0.749,
    "args": {
            "cid": 57, "correlation": 1526
    }
  },
  {
    "ph": "f", "id": 1526, "pid": 49625, "tid": 49625, "ts": 7253740960994.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960996.299, "dur": 0.678,
    "args": {
            "cid": 57, "correlation": 1527
    }
  },
  {
    "ph": "f", "id": 1527, "pid": 49625, "tid": 49625, "ts": 7253740960996.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740960998.188, "dur": 0.806,
    "args": {
            "cid": 57, "correlation": 1528
    }
  },
  {
    "ph": "f", "id": 1528, "pid": 49625, "tid": 49625, "ts": 7253740960998.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961000.720, "dur": 0.679,
    "args": {
            "cid": 57, "correlation": 1529
    }
  },
  {
    "ph": "f", "id": 1529, "pid": 49625, "tid": 49625, "ts": 7253740961000.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961002.573, "dur": 0.840,
    "args": {
            "cid": 57, "correlation": 1530
    }
  },
  {
    "ph": "f", "id": 1530, "pid": 49625, "tid": 49625, "ts": 7253740961002.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961004.614, "dur": 0.565,
    "args": {
            "cid": 57, "correlation": 1531
    }
  },
  {
    "ph": "f", "id": 1531, "pid": 49625, "tid": 49625, "ts": 7253740961004.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961006.349, "dur": 0.699,
    "args": {
            "cid": 57, "correlation": 1532
    }
  },
  {
    "ph": "f", "id": 1532, "pid": 49625, "tid": 49625, "ts": 7253740961006.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961008.314, "dur": 0.693,
    "args": {
            "cid": 57, "correlation": 1533
    }
  },
  {
    "ph": "f", "id": 1533, "pid": 49625, "tid": 49625, "ts": 7253740961008.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961010.267, "dur": 0.633,
    "args": {
            "cid": 57, "correlation": 1534
    }
  },
  {
    "ph": "f", "id": 1534, "pid": 49625, "tid": 49625, "ts": 7253740961010.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961012.295, "dur": 0.594,
    "args": {
            "cid": 57, "correlation": 1535
    }
  },
  {
    "ph": "f", "id": 1535, "pid": 49625, "tid": 49625, "ts": 7253740961012.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961014.057, "dur": 0.660,
    "args": {
            "cid": 57, "correlation": 1536
    }
  },
  {
    "ph": "f", "id": 1536, "pid": 49625, "tid": 49625, "ts": 7253740961014.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961015.974, "dur": 0.577,
    "args": {
            "cid": 57, "correlation": 1537
    }
  },
  {
    "ph": "f", "id": 1537, "pid": 49625, "tid": 49625, "ts": 7253740961015.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961017.799, "dur": 0.662,
    "args": {
            "cid": 57, "correlation": 1538
    }
  },
  {
    "ph": "f", "id": 1538, "pid": 49625, "tid": 49625, "ts": 7253740961017.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961019.807, "dur": 0.831,
    "args": {
            "cid": 57, "correlation": 1539
    }
  },
  {
    "ph": "f", "id": 1539, "pid": 49625, "tid": 49625, "ts": 7253740961019.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961021.993, "dur": 0.751,
    "args": {
            "cid": 57, "correlation": 1540
    }
  },
  {
    "ph": "f", "id": 1540, "pid": 49625, "tid": 49625, "ts": 7253740961021.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961025.871, "dur": 0.795,
    "args": {
            "cid": 57, "correlation": 1541
    }
  },
  {
    "ph": "f", "id": 1541, "pid": 49625, "tid": 49625, "ts": 7253740961025.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961028.005, "dur": 0.732,
    "args": {
            "cid": 57, "correlation": 1542
    }
  },
  {
    "ph": "f", "id": 1542, "pid": 49625, "tid": 49625, "ts": 7253740961028.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961029.900, "dur": 0.723,
    "args": {
            "cid": 57, "correlation": 1543
    }
  },
  {
    "ph": "f", "id": 1543, "pid": 49625, "tid": 49625, "ts": 7253740961029.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961031.812, "dur": 0.681,
    "args": {
            "cid": 57, "correlation": 1544
    }
  },
  {
    "ph": "f", "id": 1544, "pid": 49625, "tid": 49625, "ts": 7253740961031.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961033.742, "dur": 0.717,
    "args": {
            "cid": 57, "correlation": 1545
    }
  },
  {
    "ph": "f", "id": 1545, "pid": 49625, "tid": 49625, "ts": 7253740961033.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961035.674, "dur": 0.769,
    "args": {
            "cid": 57, "correlation": 1546
    }
  },
  {
    "ph": "f", "id": 1546, "pid": 49625, "tid": 49625, "ts": 7253740961035.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961037.784, "dur": 0.814,
    "args": {
            "cid": 57, "correlation": 1547
    }
  },
  {
    "ph": "f", "id": 1547, "pid": 49625, "tid": 49625, "ts": 7253740961037.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961039.905, "dur": 0.793,
    "args": {
            "cid": 57, "correlation": 1548
    }
  },
  {
    "ph": "f", "id": 1548, "pid": 49625, "tid": 49625, "ts": 7253740961039.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961041.914, "dur": 0.825,
    "args": {
            "cid": 57, "correlation": 1549
    }
  },
  {
    "ph": "f", "id": 1549, "pid": 49625, "tid": 49625, "ts": 7253740961041.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961043.934, "dur": 0.838,
    "args": {
            "cid": 57, "correlation": 1550
    }
  },
  {
    "ph": "f", "id": 1550, "pid": 49625, "tid": 49625, "ts": 7253740961043.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961045.998, "dur": 0.841,
    "args": {
            "cid": 57, "correlation": 1551
    }
  },
  {
    "ph": "f", "id": 1551, "pid": 49625, "tid": 49625, "ts": 7253740961045.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961048.004, "dur": 0.736,
    "args": {
            "cid": 57, "correlation": 1552
    }
  },
  {
    "ph": "f", "id": 1552, "pid": 49625, "tid": 49625, "ts": 7253740961048.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961050.227, "dur": 0.828,
    "args": {
            "cid": 57, "correlation": 1553
    }
  },
  {
    "ph": "f", "id": 1553, "pid": 49625, "tid": 49625, "ts": 7253740961050.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961052.539, "dur": 0.576,
    "args": {
            "cid": 57, "correlation": 1554
    }
  },
  {
    "ph": "f", "id": 1554, "pid": 49625, "tid": 49625, "ts": 7253740961052.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961054.417, "dur": 0.696,
    "args": {
            "cid": 57, "correlation": 1555
    }
  },
  {
    "ph": "f", "id": 1555, "pid": 49625, "tid": 49625, "ts": 7253740961054.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961056.315, "dur": 0.725,
    "args": {
            "cid": 57, "correlation": 1556
    }
  },
  {
    "ph": "f", "id": 1556, "pid": 49625, "tid": 49625, "ts": 7253740961056.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961058.361, "dur": 0.650,
    "args": {
            "cid": 57, "correlation": 1557
    }
  },
  {
    "ph": "f", "id": 1557, "pid": 49625, "tid": 49625, "ts": 7253740961058.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961060.318, "dur": 0.794,
    "args": {
            "cid": 57, "correlation": 1558
    }
  },
  {
    "ph": "f", "id": 1558, "pid": 49625, "tid": 49625, "ts": 7253740961060.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961062.429, "dur": 0.788,
    "args": {
            "cid": 57, "correlation": 1559
    }
  },
  {
    "ph": "f", "id": 1559, "pid": 49625, "tid": 49625, "ts": 7253740961062.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961064.489, "dur": 0.805,
    "args": {
            "cid": 57, "correlation": 1560
    }
  },
  {
    "ph": "f", "id": 1560, "pid": 49625, "tid": 49625, "ts": 7253740961064.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961066.584, "dur": 0.752,
    "args": {
            "cid": 57, "correlation": 1561
    }
  },
  {
    "ph": "f", "id": 1561, "pid": 49625, "tid": 49625, "ts": 7253740961066.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961068.483, "dur": 0.799,
    "args": {
            "cid": 57, "correlation": 1562
    }
  },
  {
    "ph": "f", "id": 1562, "pid": 49625, "tid": 49625, "ts": 7253740961068.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipEventElapsedTime", "pid": 49625, "tid": 49625,
    "ts": 7253740961070.515, "dur": 0.686,
    "args": {
            "cid": 57, "correlation": 1563
    }
  },
  {
    "ph": "f", "id": 1563, "pid": 49625, "tid": 49625, "ts": 7253740961070.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 49625, "tid": 49625,
    "ts": 7253740961564.608, "dur": 2.915,
    "args": {
            "cid": 48, "correlation": 1564
    }
  },
  {
    "ph": "f", "id": 1564, "pid": 49625, "tid": 49625, "ts": 7253740961564.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958698.302, "dur": 63.946,
    "args": {
      "External id": 2836,      "device": 2, "stream": 0,      "correlation": 1428, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1428, "pid": 2, "tid": 0, "ts": 7253740958698.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740958762.249, "dur": 75.773,
    "args": {
      "External id": 2837,      "device": 2, "stream": 0,      "correlation": 1429, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1429, "pid": 2, "tid": 0, "ts": 7253740958762.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740958838.023, "dur": 63.144,
    "args": {
      "External id": 2838,      "device": 2, "stream": 0,      "correlation": 1430, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1430, "pid": 2, "tid": 0, "ts": 7253740958838.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740958901.168, "dur": 74.210,
    "args": {
      "External id": 2840,      "device": 2, "stream": 0,      "correlation": 1431, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1431, "pid": 2, "tid": 0, "ts": 7253740958901.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740958975.379, "dur": 80.023,
    "args": {
      "External id": 2841,      "device": 2, "stream": 0,      "correlation": 1432, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1432, "pid": 2, "tid": 0, "ts": 7253740958975.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959062.619, "dur": 45.784,
    "args": {
      "External id": 2843,      "device": 2, "stream": 0,      "correlation": 1433, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1433, "pid": 2, "tid": 0, "ts": 7253740959062.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740959108.404, "dur": 77.297,
    "args": {
      "External id": 2846,      "device": 2, "stream": 0,      "correlation": 1434, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1736.3674092473286
    }
  },
  {
    "ph": "f", "id": 1434, "pid": 2, "tid": 0, "ts": 7253740959108.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959193.680, "dur": 5.732,
    "args": {
      "External id": 2853,      "device": 2, "stream": 0,      "correlation": 1435, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1435, "pid": 2, "tid": 0, "ts": 7253740959193.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959199.413, "dur": 6.093,
    "args": {
      "External id": 2854,      "device": 2, "stream": 0,      "correlation": 1436, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1436, "pid": 2, "tid": 0, "ts": 7253740959199.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740959205.507, "dur": 80.504,
    "args": {
      "External id": 2855,      "device": 2, "stream": 0,      "correlation": 1437, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1437, "pid": 2, "tid": 0, "ts": 7253740959205.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959286.012, "dur": 65.069,
    "args": {
      "External id": 2856,      "device": 2, "stream": 0,      "correlation": 1438, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1438, "pid": 2, "tid": 0, "ts": 7253740959286.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740959351.082, "dur": 77.377,
    "args": {
      "External id": 2857,      "device": 2, "stream": 0,      "correlation": 1439, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1439, "pid": 2, "tid": 0, "ts": 7253740959351.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959428.460, "dur": 62.422,
    "args": {
      "External id": 2858,      "device": 2, "stream": 0,      "correlation": 1440, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1440, "pid": 2, "tid": 0, "ts": 7253740959428.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740959490.883, "dur": 74.450,
    "args": {
      "External id": 2860,      "device": 2, "stream": 0,      "correlation": 1441, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1441, "pid": 2, "tid": 0, "ts": 7253740959490.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740959565.334, "dur": 80.664,
    "args": {
      "External id": 2861,      "device": 2, "stream": 0,      "correlation": 1442, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1442, "pid": 2, "tid": 0, "ts": 7253740959565.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959653.216, "dur": 45.423,
    "args": {
      "External id": 2863,      "device": 2, "stream": 0,      "correlation": 1443, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1443, "pid": 2, "tid": 0, "ts": 7253740959653.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740959698.640, "dur": 81.025,
    "args": {
      "External id": 2866,      "device": 2, "stream": 0,      "correlation": 1444, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1656.4772788981315
    }
  },
  {
    "ph": "f", "id": 1444, "pid": 2, "tid": 0, "ts": 7253740959698.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959786.883, "dur": 5.612,
    "args": {
      "External id": 2873,      "device": 2, "stream": 0,      "correlation": 1445, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1445, "pid": 2, "tid": 0, "ts": 7253740959786.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959792.496, "dur": 5.571,
    "args": {
      "External id": 2874,      "device": 2, "stream": 0,      "correlation": 1446, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1446, "pid": 2, "tid": 0, "ts": 7253740959792.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740959798.068, "dur": 80.344,
    "args": {
      "External id": 2875,      "device": 2, "stream": 0,      "correlation": 1447, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1447, "pid": 2, "tid": 0, "ts": 7253740959798.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740959878.413, "dur": 63.745,
    "args": {
      "External id": 2876,      "device": 2, "stream": 0,      "correlation": 1448, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1448, "pid": 2, "tid": 0, "ts": 7253740959878.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740959942.159, "dur": 74.851,
    "args": {
      "External id": 2877,      "device": 2, "stream": 0,      "correlation": 1449, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1449, "pid": 2, "tid": 0, "ts": 7253740959942.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960017.011, "dur": 63.064,
    "args": {
      "External id": 2878,      "device": 2, "stream": 0,      "correlation": 1450, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1450, "pid": 2, "tid": 0, "ts": 7253740960017.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740960080.076, "dur": 74.811,
    "args": {
      "External id": 2880,      "device": 2, "stream": 0,      "correlation": 1451, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1451, "pid": 2, "tid": 0, "ts": 7253740960080.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740960154.888, "dur": 81.025,
    "args": {
      "External id": 2881,      "device": 2, "stream": 0,      "correlation": 1452, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1452, "pid": 2, "tid": 0, "ts": 7253740960154.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960242.008, "dur": 44.582,
    "args": {
      "External id": 2883,      "device": 2, "stream": 0,      "correlation": 1453, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1453, "pid": 2, "tid": 0, "ts": 7253740960242.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7253740960286.591, "dur": 77.938,
    "args": {
      "External id": 2886,      "device": 2, "stream": 0,      "correlation": 1454, "kind": "DtoD",      "bytes": 134217728, "memory bandwidth (GB/s)": 1722.0868628029614
    }
  },
  {
    "ph": "f", "id": 1454, "pid": 2, "tid": 0, "ts": 7253740960286.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::reciprocal_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960370.463, "dur": 5.372,
    "args": {
      "External id": 2893,      "device": 2, "stream": 0,      "correlation": 1455, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1455, "pid": 2, "tid": 0, "ts": 7253740960370.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960375.836, "dur": 5.211,
    "args": {
      "External id": 2894,      "device": 2, "stream": 0,      "correlation": 1456, "kind": "Dispatch Kernel",      "grid":     [4, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1456, "pid": 2, "tid": 0, "ts": 7253740960375.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740960381.048, "dur": 80.904,
    "args": {
      "External id": 2895,      "device": 2, "stream": 0,      "correlation": 1457, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1457, "pid": 2, "tid": 0, "ts": 7253740960381.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960461.953, "dur": 62.142,
    "args": {
      "External id": 2896,      "device": 2, "stream": 0,      "correlation": 1458, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1458, "pid": 2, "tid": 0, "ts": 7253740960461.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740960524.096, "dur": 76.735,
    "args": {
      "External id": 2897,      "device": 2, "stream": 0,      "correlation": 1459, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1459, "pid": 2, "tid": 0, "ts": 7253740960524.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7253740960600.832, "dur": 63.585,
    "args": {
      "External id": 2898,      "device": 2, "stream": 0,      "correlation": 1460, "kind": "Dispatch Kernel",      "grid":     [32768, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1460, "pid": 2, "tid": 0, "ts": 7253740960600.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7253740960664.418, "dur": 74.691,
    "args": {
      "External id": 2900,      "device": 2, "stream": 0,      "correlation": 1461, "kind": "Dispatch Kernel",      "grid":     [65536, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1461, "pid": 2, "tid": 0, "ts": 7253740960664.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7253740960739.110, "dur": 80.263,
    "args": {
      "External id": 2901,      "device": 2, "stream": 0,      "correlation": 1462, "kind": "Dispatch Kernel",      "grid":     [131072, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1462, "pid": 2, "tid": 0, "ts": 7253740960739.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 0,
    "args": {
      "sort_index": 49625
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7253740836738.070, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7253740836738.070, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7253740836738.070, "pid": 0, "tid": 0,
    "args": {
      "name": "thread 0 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7253740836738.070, "pid": 2, "tid": 0,
    "args": {
      "name": "stream 0 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 49625,
    "args": {
      "name": "thread 49625 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 49625,
    "args": {
      "sort_index": 49625
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 49625,
    "args": {
      "name": "thread 49625 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7253740836738.070, "pid": 49625, "tid": 49625,
    "args": {
      "sort_index": 49625
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 7253740836602.630, "dur": 124981.821,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7253740836602.630,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 7253740836602.630
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 7253740976576.020
  }
  ],
  "traceName": "trace_quark_eager_no_autograd.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1743521598000000000
}